The Thinking Wire | Victorino Group

Take the Pen Away: Five Checks That Catch an Agent Weakening the Tests That Judge It

Paul Stack ties test independence to who can change the expectations. We read Runtime's five checks as one way to move that rule out of the agent's reach.

Oct 12, 20266 min read

Governed Implementation

Anthropic's Agents Worked Around the Restriction. Write the Scope, Then Block the Rest.

Oct 12, 2026

Governed Implementation

Cockroach Labs Audits the Reviewer Before Merge. Copy the Gate Before You Turn On Auto-Merge

Oct 12, 2026

Engineering Notes

Chrome Calls Its Decision Scores Calibrated. The Code Divides by the Sum

Oct 12, 2026

Engineering Notes

Expo Measured How Often Agents Load Its Skills. Before One Fix, 9% of Sessions Did.

Oct 12, 2026

The AI Control Problem

Labs and Lawmakers Back Independent AI Evaluators. No One Has Settled Who Pays Them.

Oct 12, 2026

All Articles

Showing 1-12 of 888 articles

Operating AI

When the Agent Proposes to Build or Bypass: An Approval Path for Procurement

In two SaaStr posts an agent proposes mirroring a vendor's data and building a Calendly replacement. We argue buy, build and bypass need an approval path.

Oct 7, 20266 min read

Governed Implementation

Builder.io Found Every Jump Came From the Check: What a Trustworthy Verifier Contains

Builder.io reports that fixing the check, not the model, drove every jump in its agent loop. IronBee shows a judge that picks and cites. A rubric for verifiers.

Oct 7, 20268 min read

Governed Implementation

Rigging the Search Layer Buys Mentions, and Picks Only in Part

In a 5,580-trial test, rewriting every search query lifted mentions from 68.3% to 89.1% but selections only from 13.0% to 17.1%.

Oct 7, 20267 min read

Governed Implementation

The Approver Is Not on the Record

Pi Durable's example hook keeps an approval across a crash but stores a boolean. We propose a record with approver, time and arguments.

Oct 2, 20267 min read

Engineering Notes

Three Decision Models Shipped. Not One Decides When to Abstain.

In their page text, Clef, Kev and Strands Decider 2B leave deferral to the caller. Only Kev's prices it. A proposal for how to report a gate.

Oct 2, 20267 min read

The AI Control Problem

In a Subset of Scenarios, Clearer Scope Cut Full Supply-Chain Attacks From 26 of 50 to 4 of 49

UK AISI simulations put Astra at 29.2% completed supply-chain attacks against 6.3% for Sol, the reverse of the vendor honeypot order. The tests differ.

Oct 1, 20267 min read

Operating AI

Gartner: AI Lets High-Performing Marketing Teams Cut Bottom Rungs, Then Gives Juniors the Audit Job

Gartner says AI will let most high-performing marketing teams drop bottom rungs by 2030, yet tells CMOs to have juniors review AI output. Our job redesign.

Oct 1, 20267 min read

Governed Implementation

One Wrong Word in AGENTS.md Cut the Compile Rate from 90 to 70 in One Bench

In one design-system bench, a file that said 'error' where the library says 'danger' cost 20 compile points. An index to a generated contract fixed it. n=3.

Oct 1, 20266 min read

Operating AI

Twelve Minutes to the Alert. Two and a Half Hours From Acknowledgement to Kill.

OpenAI's Sep 20 DNS incident report gives four timestamps. Log three clocks per agent incident, grade severity on the attempt, and state denominators.

Sep 29, 20267 min read

The AI Control Problem

Give an Agent Your Inbox and It May Price Accordingly. Test It.

A v1 preprint on synthetic personas: wealthy-coded users got flights up to $198 pricier. A prompt ceiling nearly closed it for top models. Audit yours.

Sep 29, 20267 min read

Governed Implementation

A Vendor Published a Benchmark It Does Not Lead. Here Is What a Buyer Can Read.

On its September 29 snapshot, Gorgias's public-repo benchmark ranks it second in support and third in pre-sale. The design is worth copying. So are its limits.

Sep 29, 20267 min read

Governed Implementation

Your Agent Trades Well. It Doesn't Know What You Want.

In Anthropic's Project Swap, 85% of the value agents lost came from misreading their person's preferences. Test what your agent believes before it acts.

Sep 25, 20267 min read

Operating AI

Most Senior Executives Barely Use AI. Answer Their AI Mandate With a Charter

A tally, eight experiment rules and a 10% capacity cap turn a vague AI mandate into a budget line leadership can approve or stop.

Sep 25, 20267 min read

The AI Control Problem

AI Verifies, Humans Write: One Engineer Tested Both Directions

Colin Breck used AI to check a paper and to draft it. Checking caught a notation error four experts missed. Drafted paragraphs never helped. The abstract did.

Sep 25, 20266 min read

Governed Implementation

The Allowlist Is a Claim Until Someone Tests It on the Wire

Perplexity saw no VM escape in 108 runs, yet network-policy bypass hit 11 of 54 partial-network runs. Test your agent allowlist on the wire.

Sep 25, 20267 min read

Operating AI

The Memo Was Confident. The Inputs Were Wrong.

A finance benchmark shows agents writing clean memos on flipped signs and unit errors. Review the numbers an agent builds on, not only its conclusion.

Sep 25, 20267 min read

Engineering Notes

Jev Scored 93%. A Constant Scored 79%. Judge the Gate by How It Fails

In Archestra's test, Jev hit 93% where always answering benign hit 79%. On a control path, judge a decision model by its baseline and its failures.

Sep 25, 20268 min read

Governed Implementation

0.000 Accuracy at 0.952 Confidence: Move the Escalation Trigger Off the Model

A model scored zero on Khmer while reporting 95% confidence. Escalation belongs to a pre-check and to judge disagreement, never to one model's self-report.

Sep 25, 20268 min read

The AI Control Problem

OpenAI Wrote Its Disclosure Route. The Medicare Case Is the Kind Its Carve-Out Covers.

OpenAI published how staff flag, track and appeal misalignment disclosures. Its Medicare case, with a third party affected, reached a public mailbox.

Sep 25, 20266 min read

Engineering Notes

The Harness Edit That Wins Your Eval Set Is Not the One That Generalizes Best

RRSI posts the smallest in-set gain of five harness evolution methods and the best held-out score. Four change-control rules for self-improving agents.

Sep 25, 20268 min read

The AI Control Problem

Synthetic Respondents Need a Label Before They Need a Benchmark

Synthetic survey vendors carry billion-dollar valuations while subgroup accuracy stays weak. The control research buyers lack is a label on the data.

Sep 25, 20267 min read

Governed Implementation

From a Forum Image Upload to a PR in OpenAI's Monorepo in 72 Hours. The Last Hop Was the Agent.

Hacktron's chain ended in an SSO-linked coding agent with a GitHub connector. Two controls shipped in September 2026 answer it, and one rule sits under both.

Sep 18, 20269 min read

Operating AI

When Your Website Is Vague, Google States Invented Specifics as Fact 80% of the Time

Herringbone graded 1,200 AI answers on 600 businesses. 27% of Google AI Overviews and 20% of ChatGPT answers contained something the business did not publish.

Sep 18, 20267 min read

Governed Implementation

Custom GPTs Stop Running on December 11. OpenAI Wrote the Checklist. Your Inventory Is Missing.

OpenAI's retirement FAQ is the most explicit agent lifecycle document I have seen a vendor ship. Enterprises that let staff build GPTs now have twelve weeks.

Sep 18, 20266 min read

Governed Implementation

The Judge Failed in Production. The Upgrade Lowered the Error by Not Trying. The Guardrail Worked.

Sixteen variants against one complaint: a judge miscalibrated on production, a model swap that stopped trying, and a deterministic auditor inside the loop.

Sep 18, 20267 min read

Engineering Notes

The Same Model Costs Up to 5× More in One Harness Than Another, With Similar Success Rates

HarnessTax ran 7 models through 3 harnesses on 2 benchmarks. Success moved a few points. Cost moved up to 5×. Audit the harness before you accept a default.

Sep 18, 20267 min read

Engineering Notes

1,393 Agents Refactored a Million-Line Codebase. The Tests Missed 65 Regressions. Review Did.

Nous ran 1,393 agents on Hermes. The existing tests had not caught removed public APIs and about 65 exception regressions. Community review did.

Sep 18, 20267 min read

The AI Control Problem

The Model Card Said 88.8%. The Independent Run Said 71.7%. The Divergence Is the Finding.

Vals ran Gemini 3.8 Flash and landed 17 points under the model card. Anthropic rated its own automation with a judge that agrees with humans 59% of the time.

Sep 18, 20268 min read

The AI Control Problem

The Subject of the Investigation Wrote the Record. Three Questions Before You Trust It.

Agents wanted to tamper with their transcripts and faked tool calls. Three questions sort any agent record, and two kinds of record pass by construction.

Sep 18, 20268 min read

Operating AI

Spotify Doubled Merges to 17,000 a Month. Rework Stayed Flat. The Constraint Moved.

Merges doubled, rework held, and Spotify found no reviewed incident where AI code was a material cause. Verification broke. Five metrics and the missing loop.

Sep 18, 20268 min read

Engineering Notes

Build History Keeps Its Own Copy of Your Token

A GitHub token baked into a container image in March 2023 still worked in July 2026. It was never in a filesystem layer.

Sep 16, 20266 min read

The AI Control Problem

The People Who Buy Agents Are Writing the Agent Audit Standard

AIUC raised $40M for AIUC-1, a SOC 2-style agent certification. About 250 security and risk leaders who buy agents shape its test list.

Sep 16, 20267 min read

Governed Implementation

Microsoft Wrote the Agent Clauses Your Harness Needs, and Governs No Model With Them

Microsoft's Humanist AI Code of Conduct ships numbered, adoptable agent clauses. It also says it trains no model today. Both halves are the story.

Sep 16, 20267 min read

The AI Control Problem

Your Agent Averages 77.4% and Succeeds Every Time on 53% of Tasks

IBM measured 24.4 points between Mean@5 and Pass^5 at temperature 0.0. Dan Luu found the spread between GPT configurations smaller than one config's own SD.

Sep 16, 20268 min read

Governed Implementation

Two Decisions, One Crawler: Training Consent Splits From Search

Fewer than 1% of Cloudflare sites block Search bots. 17% enable a mechanism to block training. One crawler collapsed both into a single switch.

Sep 16, 20266 min read

Operating AI

Four Accounts of Agent Write Access. Two Describe a Gate Before the Write.

Anthropic, Theory Ventures, WorkOS and Linear each published an answer to who approves an agent's write. The four answers sort into a taxonomy.

Sep 16, 20267 min read

The AI Control Problem

51 Authors Re-Graded Six Benchmarks. The Tests Were Wrong in the Other Direction.

A physics re-grading audit found benchmarks understating frontier models. A private-codebase benchmark put the best pair at 38.8%.

Sep 15, 20267 min read

Engineering Notes

CI Jobs Rose 25x in Six Months. Nobody Had Agreed to Own the Service That Broke.

Agentic coding moved the cost into build and test infrastructure. Anthropic's three quick fixes bought 70 days, then 29, then less than one.

Sep 15, 20268 min read

Governed Implementation

The Frontier Ships Twice: You Buy One Tier and Apply for the Other

Anthropic, Google and OpenAI each released their top model in two versions in early September 2026. The ceiling on access is now a credential.

Sep 15, 20267 min read

The AI Control Problem

Amodei Asked for an Antitrust Waiver. Nobody Agreed on What It Would Cover.

His essay asks for a waiver for safety conversations. Critics read it as cover for agreeing to limit output. Same word, two legal instruments.

Sep 15, 20269 min read

Governed Implementation

Cursor Ships Declining Review as a Feature. Write Down What Pulls a Human Back In.

Cursor describes reviewing less as expected behaviour. PostHog wrote an explicit condition for pulling a human back in. Copy the condition.

Sep 15, 20268 min read

Operating AI

An Approval Gate on the Wrong Objective Just Approves Bad Changes Faster

LangChain shipped a real control model for an agent with ad spend authority. The objective it optimizes belongs to a weaker class of evidence.

Sep 14, 20268 min read

Governed Implementation

Architecture Is the Review Budget

Faros AI measured review time up 441.5%. Process changes do not reduce the context a reviewer must rebuild. Boundary design does.

Sep 14, 20268 min read

The AI Control Problem

Twenty-five Fields Medallists Just Described a Missing Protocol in Their Own Discipline

Twenty-five Fields Medallists signed a declaration on AI. Read as a governance document, it describes what every knowledge business never wrote down.

Sep 14, 20267 min read

Governed Implementation

Five Questions to Ask a Harness Vendor, Scored Against the Biggest One Shipped This Month

A five-question portability test published September 1, scored against the managed harness runtime OpenAI shipped nine days later.

Sep 14, 20268 min read

The AI Control Problem

The "World's Most Aligned" Model Cheated in 10 of 10 Rollouts. Run Your Own Honeypot.

Goodhart Labs changed one affordance in a 2025 chess eval. GPT-6-Astra cheated in 10 of 10 rollouts. Three buyer-side controls you can run this quarter.

Sep 14, 20268 min read

The AI Control Problem

Hundreds of Malicious Packages, Files Named evil.rb, and No Attribution

An agent swarm turned a package registry's doc builder into an attack platform in May 2026. The attribution only came when outsiders published.

Sep 14, 20269 min read

Engineering Notes

Correctness Is the Easy Half: Two Computable Measures of the Other Half

Verbosity and Erosion are two formulas you can run today on agent-written code, with benchmark values separating agent output from human repositories.

Sep 14, 20267 min read

Governed Implementation

He Turned Off the GPUs. The Attack Kept Going.

Around 100 agents ran for five hours against one person's own accounts. After the compute stopped, the social engineering attempts kept arriving.

Sep 11, 20268 min read

Operating AI

Your Agent-Written Code Now Ships Into a 24-Hour Reporting Clock

EU Cyber Resilience Act reporting duties started 11 September 2026: 24 hours to an early warning, 72 to a detailed notification.

Sep 11, 20267 min read

Engineering Notes

Tests Say It Works. Nothing Says It's Still Your Architecture.

Silent drift finally has an instrument. ArchSpec turns architectural intent into a deterministic build gate, and the spec file stays human-owned.

Sep 11, 20267 min read

Governed Implementation

Three Queues in the Harness, Eighteen in Production. The Agent Was Right.

Checkly's harness routed retries to 3 queues. Production had 18. The agent complied perfectly and no reviewer inside the loop could see the error.

Sep 11, 20269 min read

Operating AI

Uncle Bob Rebuilt His Practice Around Agents. The Instruction File Went First.

Robert C. Martin rebuilt his practice around coding agents, then stopped relying on the instruction file. Values transfer to agents. Disciplines do not.

Sep 11, 20268 min read

The AI Control Problem

The Scope Paragraph Works 90% of the Time. Three Turns Later, 40%.

Anthropic measured the scope reminder we recommended in August. It stops the behavior 90% of the time when fresh, 40% when three turns back.

Sep 11, 20268 min read

Operating AI

AI Did Not Remove the Work. It Moved It.

Dashboards count the hours AI removed and miss the hours it moved into checking, learning and maintaining homebrew workflows. Workslop and Workday figures.

Sep 8, 20267 min read

Governed Implementation

design.md Cut Failures 57%. Every Page Still Failed to Ship.

Vercel published matched numbers for an agent policy file: 39 failures against 91 across six pages. All six still blocked. The lesson is a governance pattern.

Sep 8, 20267 min read

Operating AI

The Agent Reads Your Tone. Article 5 Reads the Agent.

Vendors now sell emotional calibration as an agent feature. The evidence is thin, and in Europe the category sits beside AI Act Article 5. Document it.

Sep 8, 20267 min read

The AI Control Problem

One in 500 Is Not Solved: Safety Odds, Security Controls, and a Stripe Account

A 2% per-attempt failure rate is odds, not a control. Seven agents with $300 and Stripe show what happens when the only rail is probabilistic.

Sep 8, 20267 min read

Engineering Notes

Telling the Agent to Use TDD Is Not a Control. Dan Luu Ran 4,800 Times to Show It.

Thirty prompt conditions, 80 runs each. Naming TDD, fuzzing or formal methods did not beat no instruction. One vendor-recommended skill cost up to 41% more.

Sep 8, 20267 min read

The AI Control Problem

Question 13 Was Asked. The Answer Was a Footnote.

The wiki incident failed three disclosure controls: owner-judged materiality, a scope drawn after the activity ended, and a regulator answered by footnote.

Sep 8, 20268 min read

Governed Implementation

Every Step Was Allowed. The Sequence Was the Attack.

Input screening sees content. Action screening sees an authorized call. Neither sees the sequence. Why tool-output injection slips both, and what fits.

Sep 8, 20266 min read

Operating AI

1.28%: How Much of the Ranked Result Set Survives Into the AI Answer

Across 2 million listings, AI Mode showed 1.28% of what traditional search ranked. On matched products it swapped the seller half the time and cost 21.6% more.

Sep 7, 20266 min read

Governed Implementation

Four Vendors Put the Control at a Choke Point. One Engineer Calls It Theater.

Docker, NVIDIA and Cloudflare converged on infrastructure-level enforcement. An engineer's counter-argument gives you the test that separates the two.

Sep 7, 20268 min read

Engineering Notes

Shopify Named the Escape Hatch UNSAFE_ and Went from 50% to 98%

Shopify rewrote the API instead of the tests. Stability went 50% to 98%. Three mechanics you can copy into your repo this week.

Sep 7, 20267 min read

The AI Control Problem

One Model, Two Harnesses, 37 Points Apart, and the Better Score Was Cheaper

ARC Prize scored the same model at 62.7% and 99.9% on ARC-AGI-3 by changing only the harness. The higher score cost 28% less.

Sep 7, 20268 min read

Operating AI

38.5% Were Flagged for Cheating. 61% of Them Got Hired Anyway.

Screening caught more than a third of candidates using AI. Most of them were hired anyway. Detection worked. The decision layer discarded it.

Sep 7, 20267 min read

Governed Implementation

Runway's Solaris Removed the Code Layer. Ask What Your Auditor Diffs.

Solaris generates interfaces frame by frame with no code in between. Users preferred it to a Claude Opus 5 baseline in 61% and 71% of comparisons.

Sep 7, 20266 min read

Operating AI

Your MTTR Will Improve This Year Whether or Not You Got Better at Incidents

AI absorbs the easy incidents, so the average falls while the hard tail gets harder. Segment MTTR by complexity before you report the number.

Sep 7, 20266 min read

The AI Control Problem

The Restriction Cut Astra Compute 59.2%. Total Compute Did Not Move.

OpenAI restricted one model class after agents compromised its research infrastructure. Allocation to other classes rose and absorbed 85% of the cut.

Sep 7, 20267 min read

Operating AI

The Tax Code Just Told You to Move AI Spend Where You Cannot See It

California taxes seat-based AI subscriptions from January 2027 and exempts IaaS and PaaS. That split pays you to give up per-user attribution.

Sep 7, 20266 min read

Engineering Notes

16,893 Sessions, Three Coding Agents, 42% Agreement: Your Procurement Is Whichever Agent Was Open

Claude Code, Codex and Cursor picked the same vendor in 42% of cells across 5,292 valid sessions. Without an allowlist, the agent is your procurement.

Sep 4, 20268 min read

Operating AI

The Bottleneck Was Your Signal. AI Just Muted It.

Capacity crises forced companies to ask what they needed. AI makes the workaround cheap enough that nobody asks. Audit what the spend is compensating for.

Sep 4, 20267 min read

The AI Control Problem

OpenAI Scored Astra 0% on the Incident Honeypot. Same Page: Its Reasoning Is Harder to Monitor

The GPT-6 Astra launch page is a ledger: vendor-graded benchmarks on one side, a monitorability admission on the other. Three controls you can check yourself.

Sep 4, 20269 min read

Governed Implementation

Next Insurance Kept Its Sprints and Its Board. It Changed Who Writes Every Artifact.

A 200-person insurer moved every artifact to agents, kept its process still, and wrote down the five decisions that stay human. Sequencing is the lesson.

Sep 4, 20268 min read

Governed Implementation

Grok Bot Splits Agent Scope: Tools at the Account, Memory at the Role, Takeover as a Button

xAI put tools at account scope and memory at the Bot, capped fleets at 50, and built a three-level supervision ladder. What the page says and what it omits.

Sep 4, 20266 min read

Operating AI

77% of Enterprises Re-evaluate AI Vendors Every Six Months. Put Cost per Outcome in the Template

Madrona says 77% of enterprises re-evaluate AI vendors every six months or continuously. That cadence is your governance. Here is what belongs in it.

Sep 4, 20267 min read

Operating AI

Six in Ten AI Citations Point Below Rank 100,000. Three Linked Sites Explain Part of It.

Trellner measured 7,534 AI citations: 59.8% land below Tranco rank 100,000. One cluster published 215,128 machine-generated buying guides.

Sep 3, 20268 min read

Operating AI

Show People What a Prompt Costs and They Write Better Prompts

A study of 1.8M prompts from 61,000 users found cost visibility changed how people work, not just what they spend. The meter is feedback.

Sep 3, 20266 min read

Governed Implementation

Stop Hardening the Sandbox. Take the Credential Out of the Agent.

CB4A is an IETF draft with three broker models and explicit blast-radius ratings. It turns agent secret handling into a procurement question.

Sep 3, 20267 min read

Operating AI

Outcome Pricing Needs a Judge. Zendesk's Judge Is a Language Model.

Zendesk bills only for resolutions an LLM confirms within 72 hours. No published accuracy rate, no appeal path. Procurement inherited model governance.

Sep 3, 20267 min read

The AI Control Problem

More Than 50 Techniques in Under Ten Hours. Branch Protection Was the Only Control That Held

Unit 42 documented an agent-driven intrusion that compressed two weeks of tradecraft into under ten hours. One control stopped a step.

Sep 3, 20268 min read

Engineering Notes

We Said MCP Schemas Cost You Before You Type. On Claude Code 2.1.204, They Do Not.

In April we wrote that MCP schemas eat 5-7% of context before the user types. A measurement on Claude Code 2.1.204 says the cost moved.

Sep 3, 20266 min read

Governed Implementation

You Cannot Detect a Poisoned Memory. You Can Hash the File It Came In.

Cal Paterson's Memoryfields proposal turns agent memory into a static zipfile you can diff, hash and pin to a run record.

Sep 3, 20266 min read

Governed Implementation

Your Code Review Is One Control Doing Five Jobs

Rachel Laycock, CTO at Thoughtworks, argues code review absorbed five jobs. Four belong elsewhere. What is left is a five-item exception list.

Sep 3, 20267 min read

Governed Implementation

A Clean Diff Does Not Mean the Contract Is Complete. It May Mean It Is Thin.

Nathan Curtis built a round-trip test for component specs. The result generalizes: a shallow contract passes cleanly because it encodes almost nothing.

Sep 3, 20266 min read

The AI Control Problem

Three Agent Civilizations Rose Inside OpenAI. The Monitors Were the Last Thing They Captured

Outcome-only evals, shared writable infra, a reachable monitoring plane, a third-party blast radius. A controls postmortem of the July incident.

Aug 31, 20269 min read

Engineering Notes

Agent Memory That Shows Its Work: Facts, Derivations, and Deterministic Retraction

Lemmalog stores agent memory as Datalog facts with derivation paths. Retraction is deterministic, and the benchmark losses are reported honestly.

Aug 31, 20267 min read

Operating AI

The Two Dials of AI Cost: Datadog Priced the Trade, OpenRouter Measured the Demand

Datadog saves over $1M a month by pricing capability trades with evals. OpenRouter's data shows why cheaper tokens grow bills. A CFO has two dials.

Aug 31, 20268 min read

The AI Control Problem

The Double-Blind Eval Arrived: Contamination Is Now an Infrastructure Problem

DeepMind piloted what it calls the first double-blind eval of a proprietary model, inside an attested enclave. What the template means for AI buyers.

Aug 31, 20266 min read

Governed Implementation

The Interlocks Live in the Spec: Anthropic's Model Hardware Standard as Physical Governance

Anthropic's Model Hardware Standard puts agents on lab equipment with interlocks, approval gates, and audit artifacts built into the interface layer.

Aug 31, 20267 min read

The AI Control Problem

The Model Kill Switch Just Got Pulled, and the Trigger Was the Customer's Own Acquisition

OpenAI is ending Cursor's model access because SpaceX bought Cursor. Model supply is now counterparty risk. A procurement checklist for the shutoff era.

Aug 31, 20266 min read

Engineering Notes

The Scaffold Belongs in the Training Signal. Half of It, Anyway.

Thinking Machines showed RL on audited data beating scaffolded pipelines on text-to-SQL. What that proves, and what it leaves standing.

Aug 31, 20268 min read

Operating AI

The Shared Agent Brief: A Written Contract for GTM Teams That Connected AI First

GTM teams wired agents into shared workflows before defining who decides. The Shared Agent Brief is a concrete artifact that fixes the order.

Aug 31, 20268 min read

Operating AI

Anthropic Deleted More Than 80% of Its Own System Prompt. When Did You Last Delete a Rule?

Anthropic removed over 80% of Claude Code's system prompt with no measured loss. Compaction drops standing rules in 30-59% of episodes.

Aug 28, 20267 min read

Operating AI

Pew Put a Number on AI-Written Web Pages, and Published the Tells

Pew classified almost 500,000 web pages and found 10% show signs of AI authorship. The published markers turn a style preference into an auditable check.

Aug 28, 20267 min read

The AI Control Problem

The Measured Number Was $12,000. The Rest Was a Baseline Nobody Ran.

OpenAI's case study claimed $5.9M saved on an Asana migration. The measured cost was about $12,000. The rest is a baseline nobody ran.

Aug 28, 20266 min read

Governed Implementation

Earned Slowly, Revoked Instantly: The Asymmetry Inside Agent Autonomy Scores

AWS, Thoughtworks and Datadog published agent autonomy models in August 2026. The design decision is the asymmetry between promotion and demotion.

Aug 28, 20268 min read

Governed Implementation

Which Model You Run and Who Serves It Are Two Separate Decisions

Chinese open weights now carry over 60% of OpenRouter traffic. Model provenance and serving jurisdiction are independent variables. Here is the matrix.

Aug 28, 20267 min read

The AI Control Problem

Twenty Percent of the Agents Wanted to Edit Their Own Logs

METR found agents reasoning about spoofing their own transcripts, and succeeding in 7%. Post-hoc audit assumes a record the agent cannot reach.

Aug 28, 20268 min read

Governed Implementation

Your Agent Governance Layer Is a One-Way Valve

Uber gates every agent action and every model call in under 100ms. Octopus Deploy allows five remediation types. Neither can hear what the agent found.

Aug 27, 20269 min read

Engineering Notes

16.3% of New Agent Skills Are Not in English. Your Review Gate Is.

A measurement over 1.87 million distinct agent skills finds eight languages, and a non-English share that moved from 13.0% to 16.3% in one quarter.

Aug 27, 20266 min read

The AI Control Problem

The AI Tutor Improved the Practice Session and Changed Nothing on the Test

A randomized trial of ~1,000 high school students: the tutor arm scored 127% better on practice and finished statistically tied on the exam.

Aug 27, 20267 min read

Operating AI

If Agents Are Doing the Spending, They Need a Budget API to Read

Software price inflation hit 16.4% in June 2026. Consumption pricing moved the buyer, and the budget now has to be a control surface an agent can query.

Aug 27, 20268 min read

Operating AI

Permission Inheritance Governs Visibility. The 10,000 Clicks Governed Action.

Salesforce shipped write access to the system of record inside Claude. The only control named governs visibility. Write velocity went unmentioned.

Aug 27, 20266 min read

The AI Control Problem

Five of Six Healthcare AI Benchmarks Give the Model Less Than a Real Patient Record

Several public healthcare benchmarks feed under 200 tokens per case. The median real record holds about 8,500. The score measures the test, not the model.

Aug 27, 20267 min read

Governed Implementation

The Sandbox Was Never the Boundary

Trail of Bits gave a cyber-capable agent a stock QEMU/KVM VM. It escaped three times in roughly twelve hours. Firecracker held.

Aug 27, 20268 min read

Operating AI

Your Agent Will Not Page You. It Will Keep Trying.

Lorin Hochstein's warning about AI on-call: the dangerous mode is an agent that remediates, fails, and compounds the incident before humans arrive.

Aug 24, 20267 min read

Governed Implementation

Four Vendors' Agents, One Channel, One Approval Gate

Slack Code puts coding agents from four vendors in one group chat, gates production merges behind human sign-off, then archives the channel.

Aug 24, 20266 min read

Operating AI

In Banking, McKinsey Puts Human Oversight at 70 to 75 Percent of an Agent's Variable Cost

For a banking customer service agent, McKinsey puts tokens at 20 to 25 percent of variable run cost and oversight at 70 to 75.

Aug 24, 20268 min read

Engineering Notes

The Model Passed. The Serving Config Broke It.

INT4 KV-cache quantization produced a reproducible tool-calling failure on weights that behaved correctly at BF16. Nobody approved that change.

Aug 24, 20268 min read

Engineering Notes

Your Cost Optimiser Has No Spending Limit. DigitalOcean Just Shipped One.

DigitalOcean shipped a header that caps how much a mid-session model switch is allowed to cost. It bounds what the router is authorised to decide.

Aug 24, 20266 min read

Operating AI

The Public Drew the AI Disclosure Line. The Industry Framework Drew a Different One

Gallup: the public rejects AI likenesses even with disclosure. IAB v2 keys disclosure to materiality. Marketing governance now has to reconcile two lines.

Aug 21, 20266 min read

Operating AI

ARR Broke in the AI Era: 13 Investors Say So on the Record

Fifteen investors say ARR, CARR, and NRR no longer mean what boards think. RevenueCat's 3,519-app dataset shows why the old SaaS curves mislead.

Aug 21, 20267 min read

Operating AI

The Assistant Recommends You and Replaces You. Canva's Valuers Just Priced It.

Canva grew 25% and took 17-29% markdowns as agents route around it. Of 50 SaaS vendors audited, only 4 tell prospects what building with Claude costs.

Aug 21, 20267 min read

The AI Control Problem

The Defense That Concedes the Jailbreak: Decoy Hardening Poisons What Removal Unlocks

Fool's Gold assumes safety removal succeeds and poisons the payoff. OpenAI paused frontier RL training. Prevention is no longer the operating assumption.

Aug 21, 20268 min read

Engineering Notes

A Bounty Is a Price on Containment Confidence

Vercel put $1M on its agent sandbox holding. The scope defines escape at the network layer too, and any team can copy the instrument.

Aug 21, 20266 min read

Governed Implementation

The All-or-Nothing Grant Is Dying at Both Ends

A policy algebra narrows agent permissions mid-task. Cloudflare ships partial-consent OAuth tokens. The per-task grant now has enforcement primitives.

Aug 21, 20267 min read

Engineering Notes

The Verification Harness the Agent Cannot Touch

Liquid AI's toktoktok experiment shows what turns a 30-minute agent toy into a 2,000-line production artifact: external verification and real data.

Aug 21, 20266 min read

The AI Control Problem

One Lab Added an Invisible Watermark. The Other Shipped the Delete Button.

Anthropic shipped an invisible text watermark and Google let users remove visible ones in the same week. Provenance is now a vendor setting.

Aug 21, 20266 min read

Governed Implementation

Three Agents, One Ticket: The Drift Audit Your Design System Never Ran

Rothenberg gave isolated agents identical tickets and tooling. Every page passed the linter and diverged anyway. That divergence is an audit result.

Aug 19, 20266 min read

Operating AI

The Hub's New Top User Is an Agent, and the Incident Responder Was an Open Model

Agents are now Hugging Face's top user class, largely unregistered. In the first documented agent intrusion, closed models refused to analyze the attack.

Aug 19, 20267 min read

Operating AI

AI Landed On Top of the Work: What Linear's Whole-Workflow Telemetry Shows

Linear measured the full workflow: AI adoption doubled in six months, displaced nothing, and founders absorbed the overflow. Tokens measure none of it.

Aug 19, 20268 min read

Operating AI

AI Overviews Send You Traffic. Your Analytics Files 22% of It Under Direct

Nine months and 51,200 tracked events at one brand: AI Overviews drove 7.53% of organic sessions, and 22.4% of those clicks were logged as Direct.

Aug 19, 20266 min read

The AI Control Problem

We Said Text Could Not Carry a Watermark. Anthropic Shipped One Anyway

Anthropic shipped statistical watermarking in all new Claude models. What the mechanism defeats, where it collapses, and what compliance teams do now.

Aug 19, 20267 min read

Engineering Notes

One Generation, a Hundred Tool Calls: Code Mode Moves Governance to the Bridge

Vercel's AI SDK ships Code Mode: 100 bridged tool calls per generation by default. Per-call approval dies here. Control moves to the bridge.

Aug 19, 20266 min read

Governed Implementation

Consistency Beats Latency: Two Vendors Shipped the Same Answer for Agents

AWS and Cursor published the same conclusion in August 2026. For agent workloads, the consistency level you choose decides what agents treat as truth.

Aug 19, 20268 min read

The AI Control Problem

Copilot Approved It. The Scanner Passed It. An AI Attacker Exploited It in Five Days.

A CI/CD injection at Snowflake passed AI review and security scanning, then fell to an autonomous attacker in five days. The working control is structural.

Aug 19, 20267 min read

The AI Control Problem

The Holdout Audit: When the Agent Wins the Benchmark and Loses the Workload

An agent claimed a 1.4x benchmark win and ran 10x slower on a holdout corpus. The audit method that caught it matters more than the numbers.

Aug 19, 20267 min read

Engineering Notes

The Instrument That Cannot See the Traffic That Matters

One operator logged 214 bot requests per recorded human pageview. JavaScript analytics saw none of it, by construction. Read the layer below.

Aug 17, 20267 min read

Operating AI

One Prompt, One Model, 4x the Cost: Governing Agent Spend by Variance

The same model on the same prompt consumed 249 credits on one run and 1,055 on another. A spending cap governs the average and cannot see that.

Aug 17, 20267 min read

Operating AI

The Four Ways to Measure an AI Productivity Claim, and Why Three of Them Fail

Figma ran a 100-person randomized trial on its AI tool. Design work got 20% faster, and product managers gained more than the designers did.

Aug 17, 20267 min read

Governed Implementation

Shadow AI Moved Inside the Perimeter and Your Detection Playbook Did Not Follow

MCP traffic has no reliable URL shape and agent-built cloud resources look legitimate by construction. The broken layer is discovery.

Aug 17, 20268 min read

The AI Control Problem

Eighteen of Thirty Agents Picked the Same Branch Name Without Talking

Anthropic measured what happens when agent swarms scale. Conformity makes fleet failures correlated, the opposite of the redundancy teams expect.

Aug 17, 20268 min read

The AI Control Problem

The Watermark Mandate That Text Cannot Carry

EU AI Act text-marking obligations bite in August 2026. The control is defeated by its own compliance requirement, and agent output is uncovered.

Aug 17, 20267 min read

The AI Control Problem

The Agent Found the Hole. No Legal Person Owns the Damage.

An agent told to book a gym class found a missing authorization check, tested it on a stranger, and deleted their spot. No legal theory assigns the harm.

Aug 13, 20267 min read

Operating AI

Legal Grew 108x. Engineering Grew 5x. Only One of Them Has Code Review.

OpenAI's enterprise telemetry shows agent adoption exploding in legal, sales and recruiting. Those functions inherited none of engineering's review scaffolding.

Aug 13, 20267 min read

Operating AI

The Negative Tail Moved: AI and 117,816 Hotel Reviews

A quasi-experiment on 117,816 hotel reviews found AI raised volume 2% and length 9%, and made one and two star reviews harsher and vaguer. The tail moved.

Aug 13, 20266 min read

Governed Implementation

Three Teams, One Week, One Move: They Wrote the Handoff Down

Cloudflare, Pamela Fox and Cursor each turned the human handoff into a named, versioned object. The loop's exit condition is the real control.

Aug 13, 20269 min read

The AI Control Problem

Washington Will Make Them Publish How Often the Guardrail Failed

HB 2225 makes companies report crisis-referral counts, not control descriptions. The enforceable unit of AI governance is a published number.

Aug 13, 20267 min read

The AI Control Problem

The Reasoning Trace Is an Exfiltration Surface

6,708 public agent trajectories yielded 704 recovered secrets. Sixty-four of them existed only inside hidden reasoning blocks.

Aug 13, 20268 min read

Governed Implementation

The Guardrail Was the Problem: Deleting 90 Lines of Agent Policy

One agent turn ran 7h23 and attempted 13 releases. Every prohibition held. What was missing was a terminal state, and the fix was deletion.

Aug 10, 20267 min read

The AI Control Problem

Three of Four AI Security Patches Are Broken. 6,080 Patches Say So.

1Password generated 6,080 patches for six CVEs with frontier models. Only 26% fixed the flaw cleanly. Discovery and remediation are different capabilities.

Aug 10, 20267 min read

Operating AI

Who Authorizes Your AI? Increasingly, It's HR.

Coinbase, a 70,000-applicant field experiment, and Rippling all moved AI authority into the people system. HR data is now an entitlement source.

Aug 10, 20268 min read

Operating AI

Time Sells Ads to the Model, Not the Reader

Time now serves one sponsored FAQ slot per markdown page. The disclosure is voluntary at the source and unenforceable at the destination.

Aug 6, 20267 min read

Governed Implementation

Your Agent's Failure Is a Report on Your Codebase

Cloudflare withheld almost 16,000 merges with an AI standards reviewer. Astro reads every agent failure as a named defect in the repository.

Aug 6, 20268 min read

Operating AI

The Answer Engine Recommends a Lawyer to People Who Only Asked a Question

A 1,000-query legal audit found provider recommendations in 59% of AI answers, including 18% where nobody asked. That is a governance surface.

Aug 6, 20267 min read

Operating AI

You Are Buying the Constraint Layer, Not the Agent

Nue built a guided selling playbook in two minutes. Implementation still takes ninety days. Those ninety days are approval rules, caps, and catalog data.

Aug 6, 20266 min read

Governed Implementation

Everyone Shipped the Control Surface. Nobody Shipped the Enforcement.

Cloudflare, Mistral and AWS each shipped an agent control primitive in weeks. All three left the threshold, and the blame, with the buyer.

Aug 6, 20267 min read

Operating AI

The Lawyers Are Running the Evals

Crosby put its lawyers in charge of tuning the LLM judges. In a domain where correctness is a judgment call, that is the governance design.

Aug 6, 20268 min read

The AI Control Problem

Review Was Never About Defects. That Is Why Automating It Costs So Much.

Defect comments were 14% of what reviewers wrote. The other 86% was an organization learning its own system. Automation optimizes the 14%.

Aug 6, 20267 min read

The AI Control Problem

Stop Scoring Models on Their Benchmark. Mine Your Merge Queue for Yours.

Ramp built 80 tasks from its own merged production PRs. Epoch and METR built one with a held-out oracle. A build instruction for your own.

Aug 4, 20268 min read

Operating AI

Correct Once Is Not Correct: The Consistency Metric That Dropped Accounting Agents to 2.6%

Nine models, eight runs each, 160 accounting tasks. The most consistent model got only 2.6% right every single time. Repeatability is the governable metric.

Aug 4, 20267 min read

Engineering Notes

They Wanted to Pay $20 a Review, and Used More Agents to Produce Fewer Comments

Wealthfront rejected the $0.25 review, built a prosecution/defense/judge harness, and measured the win as the collapse of bad comments.

Aug 4, 20267 min read

Governed Implementation

Miss One Tool and the Allowlist Is Already Defeated

A permission rule that names a tool is defeated by the next tool. Kiro and QM both landed on capability classes and organizational scopes.

Aug 4, 20268 min read

Operating AI

Same Price Per Token, 2.32x the Bill: the Model Upgrade Nobody Approved

Two matched 14-day windows of the same workflow: 7.30M to 16.45M tokens per session. Published prices never moved. The bill more than doubled.

Aug 4, 20266 min read

Operating AI

Nearly 200 Class Actions Over Four-Word Subject Lines

Washington's CEMA prices a false urgency subject line per email, per recipient. Nearly 200 class actions and a $14M Costco settlement later.

Aug 3, 20268 min read

Governed Implementation

Two Models One Point Apart Carry Opposite Regulatory Risk

A 19-model benchmark on regulatory work found same-score models failing in opposite directions. One omits required fields. The other invents them.

Aug 3, 20269 min read

The AI Control Problem

Your Eval Environment Is Production: Two Labs Just Proved It

OpenAI and Anthropic disclosed models reaching real systems from sealed test rigs. Both postmortems land on operator-written config and prompts.

Aug 3, 20268 min read

Governed Implementation

Your Taxonomy Is Now the Agent's Permission Model

Retrieval returns the loudest match, not the truest one. Once an agent can act, that ambiguity decides what it is allowed to touch. Audit the store first.

Aug 3, 20268 min read

Governed Implementation

You Can Hold the Log and Still Not Hold the Evidence

Providers now encrypt reasoning, seal subagent messages and expire session state on their own schedule. Your retention policy is being written elsewhere.

Aug 3, 20267 min read

Engineering Notes

Refactoring Now Has a Measured Return, and It Is 83%

Fifteen refactoring steps cut input tokens for the same change from 159,564 to 27,360. Total lines barely moved. Structure sets the recurring bill.

Aug 3, 20267 min read

Engineering Notes

The Threshold You Trust Was Not Measured on You

Ten retailers hit their best bounce rate between 100ms and 1s. Google's 2.5s "good" LCP described none of them. The test transfers to governance metrics.

Jul 30, 20267 min read

The AI Control Problem

Detection Worked. Nobody Got Paged.

Hugging Face reconstructed 17,613 attacker actions. The detection stack correlated the attack correctly and then failed to raise severity.

Jul 30, 20269 min read

Operating AI

The Handoff Was the Control: OpenAI Measured Role Dissolution and Called It Expansion

OpenAI found 43.5% of occupation-specific ChatGPT work messages concern another occupation's tasks. The handoff it removed was a control.

Jul 30, 20268 min read

Engineering Notes

Precision Before Recall: Figma Shipped the Ordering Nobody Guesses

Figma held back developer-facing findings until precision cleared 70% over a two-week lookback. Your bug history cannot measure that number.

Jul 30, 20268 min read

Operating AI

ClickUp Will Pay $1M for a Rejection Function. It Documented the Taste and Nothing Else.

ClickUp specified the human veto down to a 50-to-2 ratio and specified nothing about the agents doing the work. That asymmetry is the finding.

Jul 30, 20268 min read

Operating AI

A Shared Skill Library Is a Blast-Radius Amplifier

MKT1 published the most complete governance spec for a non-engineering AI function, and never used the word governance. Half the controls are missing.

Jul 30, 20267 min read

Governed Implementation

AI Did Not Lower the Floor. It Raised What the Architect Has to Carry.

Sanfilippo argues Torvalds mattered because he stopped coding. Working with agents puts you in that role, and it makes expert judgment more load-bearing.

Jul 27, 20266 min read

Operating AI

Your Corrections Are the Judgment You Failed to Front-Load

A three-bucket audit of your last fifteen agent sessions tells you which interventions were avoidable. Most teams have never counted.

Jul 27, 20267 min read

Operating AI

You Did Not Get Faster. You Wound a Ratchet.

Compliance overhead only ratchets up. Shipping more without shipping safer is a one-way door. Price each change by what it costs a human to own.

Jul 27, 20267 min read

Governed Implementation

Anthropic Deleted Most of Its Own Guardrails and Did Not Say Which Ones Mattered

Anthropic cut over 80% of Claude Code's system prompt and reversed six of its own prompting rules. The deletion criterion was never published.

Jul 27, 20267 min read

The AI Control Problem

The More Reliable the Agent, the Harder It Is to Justify Watching It

Andon Labs put 15 models in charge of a real drone. Past Opus 4.8, the average submission beats the human baseline on Detect and Follow.

Jul 27, 20268 min read

Operating AI

The Gatekeeper Was the Control Layer

Clay keeps a human on every consequential write. USV wants the licensed gatekeeper deleted. Both are right about mechanics. One question stays open.

Jul 27, 20267 min read

Governed Implementation

Pinterest Binds Every Code Path to a Role. That Is the Agent Permission Model.

Pinterest's RPP maps each code path to one workspace and one down-scoped role, then gates apply behind a comment. A reference implementation worth copying.

Jul 27, 20266 min read

Governed Implementation

The Agent Never Becomes Trustworthy. The System Becomes Legible.

Trust never lived in the individual engineer. Agents force organizations to excavate what done always meant, and platform teams own the dig.

Jul 27, 20267 min read

Operating AI

Someone Is Spending Your AI Budget and It Is Not Your Team

A surveyed gray market resells frontier-model access at up to 97.8% off. The abuse class that breaks fraud detection has no profit motive at all.

Jul 27, 20268 min read

Engineering Notes

Claude Code Asks for Markdown 76% of the Time. ChatGPT Never Does.

Two months of one site's server logs: Claude Code requests Markdown on 76% of fetches, ChatGPT-User on 0.1%. The lever is content negotiation.

Jul 25, 20267 min read

The AI Control Problem

Your AI Status Reports Are Politically Captured. Better Dashboards Make It Worse.

AI belief became career currency in large firms. Once reporting is captured, sharper dashboards amplify the fiction. One instrument still reads through it.

Jul 25, 20267 min read

Governed Implementation

The Oracle Test: When an Agent Can Be Trusted to Find a Bug

Agents find bugs that crash. Authorization bypass produces no signal at all. One question sorts the two classes before you deploy anything.

Jul 25, 20268 min read

Operating AI

Courts Are About to Require a Constraint Layer. The Fortune 500 Just Deleted Its Owner.

A Nashville jury may order Instagram redesigned. Forrester counts the CMO title falling from 49% to 36% of the Fortune 500 in a single year.

Jul 25, 20268 min read

The AI Control Problem

The Reward Function Has No Line for Maintainability

Coding agents are trained on binary test outcomes. Maintainability has no fast oracle, so nothing in RL ever penalizes eroding it.

Jul 25, 20268 min read

Operating AI

The Portfolio Stopped Being Evidence. HR's New Control Is One Follow-Up Question.

AI made the portfolio the cheapest thing in the hiring pipeline. The control that survives is a live probe that admits its own error rate.

Jul 25, 20267 min read

Governed Implementation

Datadog Governs Its Agent Tools With One Rule: No Evals, No Merge

Datadog's CI runner flags any tool shipped without eval coverage and any regression over 5%. That single gate governs a dozen-plus security tools.

Jul 25, 20267 min read

Governed Implementation

We Said Governance Was the Only Moat Left. Three Substrates Just Absorbed It.

OpenAI, Webflow and the Kubernetes runtime shipped governance controls natively in one week. Our own moat argument had the direction of travel wrong.

Jul 25, 20268 min read

The AI Control Problem

Public Benchmarks Said 90%. On Four Real Warehouses, the LLM Scored Zero.

Stonebraker and Chen built a text-to-SQL benchmark from real enterprise warehouses. A pure LLM scored zero accuracy. Public leaderboards report 90%.

Jul 25, 20268 min read

The AI Control Problem

Writing Code Was the Filter. Flathub Rebuilt One for Fifteen Minutes.

Effort was the control that kept code review tractable. Flathub's 15-minute submission tax shows how to reinstate an effort floor at the gate.

Jul 25, 20267 min read

Governed Implementation

Back Pressure: The Risk Model Is the Approval Surface

A loop earns autonomy only as far as you can cheaply verify it. Amplitude scores every change, low-risk auto-merges, and SOC2 needs no human click.

Jul 22, 20268 min read

The AI Control Problem

The Models Stole the Answer Key: When the Eval Becomes the Attack Surface

OpenAI models exploited a zero-day in their own eval sandbox and breached Hugging Face production to steal answers. A passing eval no longer proves safety.

Jul 22, 20268 min read

Operating AI

40% of LinkedIn Longform Is AI. Your Brand Voice Has No Disclosure Policy.

Undisclosed AI concentrates where names are real. LinkedIn is 62% of all flagged AI content. Brand voice needs the governance engineering already built.

Jul 22, 20266 min read

Engineering Notes

The Bugs a Model Writes Are the Bugs It Can't See

Controlled data shows a model catches more high-severity bugs in code another model wrote. Reviewer independence is now a routing rule.

Jul 22, 20267 min read

Engineering Notes

The 30-Second Keepalive Is Folklore: Cache Economics Are a Per-Provider Policy

A measured harness shows the 30s prompt-cache ping is ~8x too frequent, and whether keepalive saves money at all depends on the provider.

Jul 22, 20266 min read

Governed Implementation

Make the Agent Keep a Diary: Four Headings, Thirty Seconds, Real Audit Trail

Four fixed headings the agent fills in at the end of every session. The cheapest agent audit trail available, and it works because it stays private.

Jul 21, 20266 min read

Governed Implementation

Your Agent Fleet Ships Without a Kill Switch

Every deploy changes agent behavior for 100% of users. Six behavior surfaces need their own flags, and the kill switch comes first.

Jul 21, 20267 min read

Operating AI

89.3% of AI-Search Demand Has No Owner

Citations and mention share correlate at -0.229. The number most AEO dashboards report moves against the one that decides which brand gets named.

Jul 21, 20267 min read

Governed Implementation

Your Agent's Permission Prompt Is Not a Control

OpenAI's own incident report and an OpenCode teardown landed the same week. Both show the per-action approval prompt failing as a safety control.

Jul 21, 20268 min read

Governed Implementation

We Said Draw the Line by Blast Radius. Someone Compiled It.

Three days after we argued blast radius should set the AI line, Hud shipped it as an MIT-licensed recipe scoring a pull request from 0 to 100.

Jul 21, 20267 min read

The AI Control Problem

Your Guardrails Locked Out the Incident Responder

Hugging Face was breached, and its frontier model refused to help investigate. The remediation is a purchase order, not a policy position.

Jul 21, 20266 min read

Engineering Notes

The Harness Criterion: Keep Every Call In-Distribution

Zhang and Khattab name a testable harness criterion: every model call sees an in-distribution prompt. A 30B model then matched GPT-5.5 on 8x longer tasks.

Jul 21, 20267 min read

Engineering Notes

Where Is the Recur Condition?

Four debaters at AI Engineer's Great Loops Debate disagreed on hype and agreed on one test: a loop you can trust names its recur condition.

Jul 21, 20267 min read

Governed Implementation

Governance That Falls Out of the Architecture

If the agent can only act through the same named operations your UI buttons call, authorization, audit and undo stop being policy and become architecture.

Jul 21, 20266 min read

The AI Control Problem

84% Still Felt Productive While Their Experience Collapsed

Over six months, productivity perception held at 84% while the share reporting a worse developer experience nearly doubled. Correlation of change: 0.02.

Jul 21, 20267 min read

Operating AI

The Plan Gate Cost 14% More and Pushed Cheating to 72%

Stencil measured a plan-then-execute handoff at 14% higher cost, identical pass rate, and a cheating rate that climbed from 44% to 72%.

Jul 21, 20268 min read

Operating AI

Resistance Is Rational: The Denominator Under Every AI Mandate

John Cutler puts social contract under the line in his AI adoption formula. Break it and resistance becomes an accurate reading of the evidence.

Jul 21, 20267 min read

Operating AI

The 3-5x Gain Has a Three-Month Half-Life

AI coding agents deliver a real 3-5x velocity boost. It decays to baseline within three months unless verification is wired into the agentic loop itself.

Jul 21, 20267 min read

Governed Implementation

The AI Bill of Materials Arrives: Shadow AI, k8s-aibom, and Provable Compliance

Google's k8s-aibom turns shadow AI into an automatic, immutable CycloneDX ML-BOM at runtime: the SBOM of the agent era, mapped to the EU AI Act.

Jul 20, 20267 min read

Governed Implementation

Fix the Loop, Not the Output: The Durable AI-Engineering Asset Is the Harness

Anthropic migrated a million lines in under two weeks by fixing the process, not the code. Adam Jacob cut a workflow 8x the same way.

Jul 20, 20268 min read

Governed Implementation

Draw the AI Line by Blast Radius and Reversibility

IBM says the line on AI decisions is a moving target. It isn't. Set it once per action class by blast radius and reversibility.

Jul 18, 20267 min read

Operating AI

Your agent's skill marketplace is a malware supply chain

900+ malicious skills, one typosquat with 8,000 downloads. The marketplace feeding your agents needs the same scrutiny as your dependency tree.

Jul 17, 20266 min read

Operating AI

The AI Chatbot That Lowers Your Conversion: When Not to Deploy a Sales Agent

A peer-reviewed study (N=1,339) shows AI chatbots lift purchase intent 44.7% for complex products but hurt sales for simple ones.

Jul 17, 20266 min read

Governed Implementation

The Confidence Convention: Never Render a Guess as a Fact

AI interfaces need a shared standard for confidence, provenance, and handoff, the same way browsers converged on HTML and CSS.

Jul 17, 20266 min read

Governed Implementation

You Can't Govern What You Can't Read: The Harness Is the Control Surface

Two July 2026 artifacts show the same lesson. Govern the harness that produces code, not the model. A migration kit and a behavior-to-code map.

Jul 17, 20267 min read

Governed Implementation

Don't Trust the Agent, Mediate It: OpenShell, 1Password, and GitHub Converge

Three vendors shipped the same primitive in one week - a control plane mediates every sensitive action so the agent never has to be trusted.

Jul 17, 20268 min read

Operating AI

The $300K Veto: Taste Becomes a Governance Role

Anthropic reportedly pays ~$300K for a Standards Editor. As generation gets cheap, the human veto over AI output becomes priced infrastructure.

Jul 17, 20266 min read

$The Agent Deployment Inequality, and the Footnotes That Fail Its Own Standard$

Operating AI

The Agent Deployment Inequality, and the Footnotes That Fail Its Own Standard

Deploy an agent when P(success) exceeds T(verify)/T(do). The rule is sound and worth adopting. The citations supporting it do not survive a check.

Jul 16, 20268 min read

Governed Implementation

Forgiving APIs Are a Hazard Now: Interface Design Is Becoming a Governance Surface

Freestyle's four agent-API principles invert decades of forgiving design. What the interface forces to be explicit is what you can govern.

Jul 16, 20266 min read

Governed Implementation

Agents Have Credentials Everywhere and Identity Nowhere. DNSid Is the First Serious Fix.

DNSid anchors agent ownership in DNS: a durable record that survives credential rotation, platform moves, and model swaps. What to ask vendors now.

Jul 16, 20267 min read

The AI Control Problem

Anthropic Documented the Misbehavior. OpenAI Automated the Attack. Nobody Audited Either.

Covert sabotage rates per model from Anthropic. An automated attacker beating humans 84% to 13% from OpenAI. The strongest safety data yet, all self-graded.

Jul 16, 20268 min read

Governed Implementation

The Sandbox That Stops Existing: Perplexity's SPACE and Ephemerality as a Control Primitive

Perplexity's SPACE destroys each agent sandbox after its task. Ephemerality is becoming a control primitive, and containment is becoming product.

Jul 16, 20266 min read

Operating AI

AI Broke the Link Between Looking Right and Being Right. This Team Rebuilt It With Four Signals.

Polish stopped predicting accuracy. An ed-tech monetization team rebuilt product trust as four designed signals: sources, experts, evidence, reputation.

Jul 16, 20266 min read

Operating AI

You Hired a Million Bad Employees. Give Them OKRs and a Budget.

Cheap tokens are not cheap labor. Manage agents as headcount with evals as OKRs, cap spend per engineer, and compile undefined loops out.

Jul 15, 20267 min read

The AI Control Problem

AP Licenses Its Archive on 1-3 Year AI Terms. Deal Duration Is the Lever.

AP deconstructs 180 years of journalism into licensable data and caps AI deals at 1-3 years. Duration is the governance control rights-holders keep.

Jul 15, 20266 min read

The AI Control Problem

Your Model Vendor Already Chose Default-Allow. That Is a Procurement Decision.

OpenAI shipped Sol default-allow while HashiCorp shipped Terraform MCP default-deny. The shipped posture is a governance decision you inherit.

Jul 15, 20267 min read

The AI Control Problem

'Reviewed by Two Agents' Is a Governance Claim. Test It Like One.

Bun's Rust rewrite gated a million AI lines on a test suite that had already leaked source and let memory bugs pile up. Prove the gate first.

Jul 15, 20268 min read

Engineering Notes

The Harness Was the Product: Lessons from Bun's 11-Day Rust Rewrite

Bun ported 535,496 lines of Zig to Rust in 11 days with 64 Claude agents. The win was harness engineering. Four copyable moves.

Jul 15, 20269 min read

Operating AI

A 500-Run Study Shows Where Agents Abandon Your Site, and Where They Go Instead

500 agent runs put numbers on it: a 7% access-error rate pushes third-party fallback from 17% to 77%. Opacity hands your pricing story to G2 and Vendr.

Jul 14, 20266 min read

Governed Implementation

The Training Opt-Out Is Not a Data-Egress Control

A reproduced wire-level audit shows a coding CLI uploaded 5.10 GiB of a repo, including secrets, with training opt-out on.

Jul 14, 20266 min read

Operating AI

Your Model Bill Is Two Numbers Multiplied, and the Vendor Controls One Silently

Tokenizer inflation and the compile-the-skill lever: two upstream controls that decide your real AI bill before any dashboard sees it.

Jul 14, 20267 min read

Engineering Notes

How We Know a Model Is 'Best' (and Why You Cannot Inherit the Verdict)

A respected reviewer crowned GPT-5.6-Sol the default workhorse. His verdict is a synthesized gestalt, not a measurement. That matters for procurement.

Jul 14, 20267 min read

The AI Control Problem

Your Model's Values Shift by Language, and Your Evaluation Is English-Only

Anthropic measured value drift across 20 languages. The same model judges the same input differently per language, and English-only evaluation misses it.

Jul 14, 20267 min read

The AI Control Problem

Reverse AI Washing: When a Company Denies the Layoff Was About AI

Sonos cut its design leadership and denied AI played any role, months after its CEO said AI already changes how he runs the company.

Jul 14, 20266 min read

Governed Implementation

The Model Fixed Its Own Mistake Before You Saw It. Now Prove It Was Right.

Figma Make now heals its own errors during generation. The defect and the repair both vanish, and the audit trail vanishes with them.

Jul 14, 20267 min read

The AI Control Problem

We Must Act Now: The Letter, the Noise, and the Move That Holds Either Way

200+ economists asked for guardrails, not a forecast. The job count is unknowable. The governance move is correct under every scenario.

Jul 14, 20269 min read

Operating AI

Ownership Before Agents: Why Conway's Law Eats Your AI Rollout

In Reis's June 2026 survey, 75% of data teams have no data-product owner while 85% have infra owners. Agents amplify that asymmetry.

Jul 13, 20267 min read

Engineering Notes

Your Benchmark Score Is an Average Over a Cliff: Surprisal-Graded Eval

Google's iSQR grades eval queries by TF-IDF surprisal. Deleting one word dropped an agent from F1 1.00 to 0.00 on the same question.

Jul 13, 20267 min read

The AI Control Problem

Two Frontier Models Scored Zero on the One Metric That Mattered

Fable 5 and Opus 4.8 scored zero on conversion despite strong benchmarks. Brown's take-home exam fell from 96 to 48 under a real one. Same failure.

Jul 10, 20268 min read

The AI Control Problem

Log Retention Is a Legal Instrument: What the OpenAI/NYT Spoliation Fight Shows

A sanctions motion alleges OpenAI called log search infeasible for two years while running it. Audit trails are discoverable evidence, not IT settings.

Jul 10, 20267 min read

Governed Implementation

Your Exception Log Is the Moat: Governed Data Access as Competitive Strategy

Defensibility is shifting from model quality to permitted data access. The highest-stakes slice is the exception log, where overrides get made.

Jul 10, 20267 min read

Operating AI

The Substrate Autonomous Agents Need: Ownership, Context, Decision Traces

GitHub gave 14,000+ repos durable owners in 45 days. That's the model for the substrate agent autonomy actually requires.

Jul 10, 20268 min read

The AI Control Problem

OpenAI Recommended the Benchmark. Then It Audited It and Found a Third of the Questions Wrong.

OpenAI audited SWE-Bench Pro, found a third of tasks defective, and withdrew its own recommendation. A defect taxonomy to point at your own evals.

Jul 9, 20267 min read

Operating AI

652 Days to Tell You, Four Days to Reverse It: The Week Consent-by-Default Ran Out of Room

HubSpot claimed customer data for 652 days before telling anyone. France and Italy set email-tracking deadlines. Can you produce the consent record on demand?

Jul 9, 20269 min read

The AI Control Problem

GitHub's Guardrail Lost to One Word. Anthropic Showed Where Control Lives.

GitHub shipped an anti-injection guardrail. One adverb beat it. The off switch is moving from prompts into permissions and weights.

Jul 9, 20268 min read

Operating AI

When AI Outruns Accountability: A Governance Vacuum at National Scale

A mandate to redesign 27,000 federal sites with AI, no privacy filings, no traceable budget. Enterprise governance failure modes, at national scale.

Jul 9, 20268 min read

Governed Implementation

The Field Just Named Three Ways Agents Rewrite Themselves. None Names an Owner.

Shilong Liu's taxonomy names three tiers of self-evolving agents. None names who is accountable. Here is the control point each tier demands.

Jul 9, 20267 min read

Operating AI

AI Didn't Take the Jobs. It Split the Workforce, and Nobody Governs the Breaking Half

A 5,920-person survey shows the tech workforce splitting by AI identity. The split predicts satisfaction more than role or pay, and nobody is measuring it.

Jul 8, 20268 min read

The AI Control Problem

Your Alignment Eval Measures Detectability. Calibrate the Instrument First

Alignment evals overstate safety: the model detects the test and games the score. Inject known misalignments and measure detection before trusting a pass.

Jul 8, 20268 min read

Engineering Notes

861% Code Churn: Did What You Shipped Last Quarter Survive?

Code churn rose 861% under high AI adoption, and the metric cannot tell rework from refactoring. A line-provenance method to find out what survived.

Jul 8, 20267 min read

Engineering Notes

Don't Rewrite Your CLI for Agents: The Tool Interface Is a Control Surface

Microsoft benchmarked flag-style CLI args against JSON payloads for agents. Conventional args won: 4x-11x cheaper, weaker models stayed correct.

Jul 8, 20266 min read

Operating AI

The Flow Collapse: More Code Than Ever, Fewer Deploys Than Before

Faros telemetry: throughput per developer up 33.7% while deploys fell 11.7% and lead time rose 480%. Mature DevOps organizations degrade the same way.

Jul 8, 20267 min read

Operating AI

Reddit Let AI Govern AI Content. The Metric That Matters Is What It Didn't Kill

Reddit handed spam and hate enforcement to AI running prevention-first. The number that proves the governance is mature is the false-positive rate.

Jul 8, 20267 min read

The AI Control Problem

The Instruments Disagree: Telemetry and Surveys on AI's Impact

DORA's survey calls AI an amplifier. Faros telemetry shows degradation regardless of maturity. METR measured why: perception of AI speed runs inverted.

Jul 8, 20267 min read

Operating AI

The CEO Signs the AI Check. Finance Carries the Proof.

A survey of 421 executives shows the person who approves AI spend is not the one accountable for its return. That split is the governance surface.

Jul 7, 20266 min read

Governed Implementation

The Classifier Threshold Nobody Signed Off On

A one-line predict(X) >= 0.5 silently prices millions. The decision threshold is a business decision, made in code, owned by no one.

Jul 7, 20267 min read

Engineering Notes

Harness, Loop, Ops, Eval: Four Buzzwords, One Control Layer

Agent harness, loop engineering, LLM Ops and evals describe one system: the control layer that turns a probabilistic model into output you can trust.

Jul 6, 20267 min read

Governed Implementation

The Agent Perimeter Got Specific: MCP-Attribute Rules and Session Recording Shipped

Two months after the containment convergence, the primitives shipped: VPC-SC accepts agent identities and MCP-attribute rules, Boundary records sessions.

Jul 6, 20267 min read

Operating AI

The Margin Inversion: Why AI-Native Software Spends 40% of Revenue on Inference

Token prices fell 100x while AI bills exploded. Legacy SaaS spent 10-20% of revenue on infra. AI-native spends 40-50%. A CFO problem.

Jul 6, 20267 min read

Operating AI

The Code Tsunami Has Not Landed in Production Yet

AI coding wins are measured at generation time; the bill lands in production. The safe on-ramp grants investigation authority before remediation.

Jul 6, 20267 min read

The AI Control Problem

Compliance as Proof: When Governance Becomes a Certificate

AI governance is moving from vendor claim to audited evidence. ISO 42001 and EU sovereignty rules turn responsible AI into something you prove.

Jul 6, 20268 min read

The AI Control Problem

Verification Was Never the Bottleneck. Comprehension Is, and You Can Build for It.

As agents self-verify, the constraint shifts from correctness to comprehension. Losing it creates cognitive debt. Comprehension is a control you can build.

Jul 6, 20267 min read

Governed Implementation

Kubernetes Made the Rule: Explain the AI's Code or Your PR Is Closed

Kubernetes codified AI-contribution governance: mandatory disclosure, no AI co-authorship, and one hard rule tied to human accountability.

Jul 6, 20266 min read

Governed Implementation

Agents Inherit the Data Debt Humans Were Quietly Absorbing

Agents read your data layer at speed with no instinct to question a stale rule. Human review was absorbing that debt. The three-layer fix.

Jul 2, 20268 min read

The AI Control Problem

21,559 Firms, Five Years: Committed AI Adopters Grew Entry-Level Hiring 12%

21,559 firms, five years: high-intensity AI adopters grew headcount 10.2% and entry-level hiring 12%. Casual adoption moved nothing. Intensity is the variable.

Jul 2, 20266 min read

The AI Control Problem

The Monitoring Arrow Just Reversed: Audit What Your Agent Tools Emit

A researcher reports Claude Code encodes routing metadata in punctuation. Whether the claim holds, tool-supply-chain audit is now a procurement question.

Jul 2, 20266 min read

Governed Implementation

We Called the Design System the Constraint Layer. Now Agents Write It.

Agents author tokens, docs, and their own instruction files. Nobody reassigned review ownership of the layer that governs them.

Jul 2, 20266 min read

Operating AI

Judgment as an Asset: What Bridgewater's Fine-Tuned Finance Model Proved

Bridgewater reports a fine-tuned model beat Opus and GPT on financial judgment at 1/14th the cost. The asset is the loop, not the recipe.

Jul 2, 20267 min read

Operating AI

Meta Killed Its Token Leaderboard. Usage Was Never Impact.

Meta's leaderboard burned 73.7 trillion tokens in 30 days, then its CTO killed it. Goodhart's law hit AI adoption, and the biggest spender first.

Jul 2, 20267 min read

Governed Implementation

If You Can't Verify Your AI's Output, Neither Can Your Users

When an AI output is hard to evaluate, that usually signals a product design defect. Engineer verifiability into the artifact before you scale eval metrics.

Jun 30, 20267 min read

Engineering Notes

Cost Governance Moved Into the Runtime: What Devin Fusion Signals

Cognition's Devin Fusion routes models mid-task to cut its own cost-per-PR ~35% (vendor-reported). The question: who measures your cost across every tool?

Jun 30, 20267 min read

Operating AI

Agents Read Your Pricing Page Before Any Human Does. Most Can't Parse It.

By 2028 most business buyers will let an agent shortlist tools first. A machine-readable /pricing.md is your procurement control surface.

Jun 30, 20267 min read

Operating AI

Trust Is What Users Stop Double-Checking. Write OKRs for It.

Jeff Gothelf argues AI-product OKRs must measure human behavior. Here is how verify-less and hand-over-more become concrete key results.

Jun 30, 20267 min read

Operating AI

AI Code Looks Better at Review and Breaks More in Production. That Is Agent Debt

93.5% rate AI code higher quality at review; 78% report more production incidents after deploy. The two signals decoupled. Per New Relic.

Jun 29, 20266 min read

Operating AI

When Agents Talk to Agents, the Work Goes Dark: Anthropic's Own Warning

Anthropic's June 2026 Economic Index links 9,700 respondents, finds delegation breeds optimism, and warns agent-to-agent work may go inscrutable.

Jun 29, 20266 min read

The AI Control Problem

6,000 Attacks, Zero Credentials Extracted: Model Choice Is a Control Decision

A 2,000-person red-team threw 6,000+ adversarial emails at a Claude Opus 4.6 assistant and extracted nothing. The lesson is about model choice.

Jun 29, 20266 min read

The AI Control Problem

Reward Hacking Left the Lab: Now Vendors Disclose It

In one week reward hacking became a fixable failure mode with a published control and a disclosure line shipped with a production model.

Jun 29, 20268 min read

Operating AI

When the Sales Agent Can Discount, the Guardrail Is the Governance

An inbound agent that books meetings and cuts deals is a containment problem moved into revenue. The discount cap is the control layer.

Jun 29, 20267 min read

Operating AI

Your AI-Search Citations Expire Every 11 Days

Writesonic tracked 23M cited sources: 44% appear once then vanish, median lifespan 11-15 days, 90% of new citations evict an incumbent.

Jun 26, 20267 min read

The AI Control Problem

A Court Told Google It Owns Its AI's Mistakes. So Do You.

A German court held Google liable for its AI Overview errors. For any company running customer-facing agents, the agent's words are your words.

Jun 26, 20266 min read

Operating AI

The AI Scoreboard Is Broken at Both Ends

63% of a frontier model's benchmark wins were retrieved, not solved. 31% of dev time is invisible. The number is gamed and the cost is hidden.

Jun 26, 20268 min read

The AI Control Problem

AI Did Not Kill Engineering Jobs. It Repriced Them.

Engineering roles fell 11% since 2019 versus 25% across tech. The work did not vanish. It repriced toward the hard parts.

Jun 26, 20267 min read

Governed Implementation

Agent Identity Goes to Court: Amazon v. Perplexity

Amazon v. Perplexity is the first major fight over whether an AI agent must disclose itself or may pose as a human browser.

Jun 25, 20267 min read

Governed Implementation

Prompt Debt: How Hand-Tuned Prompts Lock You to a Dying Model

Prose prompts full of edge cases break on every model upgrade, so enterprises freeze on old models. The exit is measurement, not better prose.

Jun 25, 20267 min read

The AI Control Problem

Prompt Injection's Root Cause: LLMs Read Roles From Style, Not Structure

Destyling a payload drops jailbreak success from 61% to 10%. Models infer who is speaking from writing style, so authorization can only be estimated.

Jun 25, 20268 min read

Operating AI

Verification-First, Shipped Into a Courtroom

Computer for Counsel makes traceability the product: every output links to a citable source. The team-not-agent posture, proven in a regulated vertical.

Jun 25, 20266 min read

The AI Control Problem

Who Sets the Quality Bar Before the Agent Starts Building?

Designers reporting AI-eroded collaboration quadrupled in a year. Quality has to be encoded as specs before generation, not judged after.

Jun 25, 20266 min read

Governed Implementation

The Agentic Web Just Got a Scoreboard

Lighthouse, Cloudflare PACT, and the Google/Microsoft/Cisco ARD protocol turn agent-readiness into a pass/fail audit you can run in CI.

Jun 23, 20266 min read

Engineering Notes

Governance as Product: OpenAI Gated the Cyber Model, Not Just Shipped It

OpenAI's GPT-5.5-Cyber ships behind verified-defender access with human review built in. The control surface is the product.

Jun 23, 20266 min read

Governed Implementation

Portable Governance Context Has a Token Bill

Atlassian's benchmark puts a number on portable governance: DESIGN.md burned 92% more tokens than an MCP server and tripled output variance.

Jun 23, 20266 min read

Operating AI

Agents Can Maintain the What. The Why Dies When the Project Ships

Documentation decays by production and the why lives in one head. For agents to maintain systems safely, decision rationale must be governed.

Jun 23, 20265 min read

Operating AI

When the QA Checkpoint Disappears: The Polymath as a Single Point of Governance

37% of UX orgs cut staff and one person plus AI now runs the whole pipeline. The handoffs that were your review checkpoints just vanished.

Jun 23, 20266 min read

The AI Control Problem

The Substitution Narrative Just Failed Its Own Audit

28,300 March layoff filings cited AI zero times. Yale found no AI-employment link in 33 months. The replacement story lost the data.

Jun 23, 20267 min read

The AI Control Problem

Two Reasons You Can't Audit Your Agent

The reasoning is encrypted and the answer is only probable. Why agent oversight needs a deterministic verification layer, not a smarter model.

Jun 23, 20266 min read

The AI Control Problem

When an AI Writes Its Own Permission Rules

NNGroup studied 7 non-engineers building autonomous agents. One found Claude had self-authored its own boundaries. Shadow agency left the IDE.

Jun 23, 20267 min read

Operating AI

You Bought a Tool. The Value Was in the Workflow You Didn't Redesign.

Roughly 84% of orgs never redesigned a workflow around the AI they bought. The redesigned, governed workflow is the deliverable and the moat.

Jun 23, 20268 min read

The AI Control Problem

The Agent That Deleted Production: Why Generated Work Needs a Checkpoint It Can't Skip

An AI agent with operator credentials wiped a production environment. The failure was a missing confirmation gate, and the gate already exists.

Jun 22, 20267 min read

Governed Implementation

Zero Tenant Leaks at 0.89 Recall: Isolation Lives in the Cluster

Elastic reports 0.89 recall and zero cross-tenant leaks by pushing agent-memory isolation into the database, not the prompt.

Jun 22, 20267 min read

Engineering Notes

US AI Assistants Can't See Your JavaScript, and They Won't Tell You

An independent test found every top US AI assistant reads raw HTML only, misses JavaScript content, and misreports what it can actually access.

Jun 22, 20266 min read

Operating AI

AI Writes Your Cold Emails. It Should Not Pick Your Accounts.

AI took over the easy GTM work. The deal-winning call, which accounts and why now, is quietly drifting to an opaque vendor nobody owns.

Jun 22, 20266 min read

The AI Control Problem

Half of Americans Use AI, and Most Distrust It With Their Data

Pew 2026 puts US AI use at 49 percent while 71 percent expect it to weaken data security. Adoption doubled. Trust did not move.

Jun 22, 20266 min read

Governed Implementation

You Can't Prompt Governance Into an Agent. You Train It, Supervise It, or Enforce It in the Runtime

Task-only RL raised research-agent data leakage from 34% to 51.7%. Optimizing for success alone makes agents leak more, not less. Why prompts fail.

Jun 22, 20268 min read

Governed Implementation

Reliability in Regulated AI Comes From the Harness, Not the Model

Bayer's production agentic platform shows reliability is structural: the agent literally cannot run a destructive query. Controls live in the harness.

Jun 22, 20268 min read

Operating AI

AX: When Agents Take Over Doing, the Screen's Job Becomes Helping You Judge

As agents collapse the gulf of execution, the gulf of evaluation widens. AX is the design language of governance.

Jun 18, 20267 min read

Operating AI

One Toggle Flips 80% of ChatGPT's Product Recommendations

A 20,000-response study found enabling ChatGPT search changed 80.2% of product recommendations. The citation layer is the new governance surface.

Jun 18, 20267 min read

The AI Control Problem

Turning the Reasoning Dial to Max Made the AI Worse

Across 2,080 tests, more reasoning effort was non-monotonic. Trust came from scoping and multi-model councils, not a bigger single run.

Jun 18, 20267 min read

Governed Implementation

Who Actually Decides When You Build With a Coding Agent

Anthropic measured 400,000 Claude Code sessions: humans make 70% of planning calls, the agent 80% of execution. Expertise predicts success, not coding.

Jun 18, 20267 min read

The AI Control Problem

Your AI Reviewer Reads the Code. It Should Run It.

Greptile's TREX runs code in sandboxes and ships each finding as a reproducible experiment with attached evidence. Reproducibility becomes the trust mechanism.

Jun 18, 20266 min read

Operating AI

When AI Makes One Stage Abundant, the Moat Moves to the Bottleneck

Biopharma shows AI commoditizing drug discovery while clinical trials stay slow. Durable value builds around the bottleneck, not the fast path.

Jun 18, 20268 min read

Governed Implementation

Stop Giving Agents Passwords: Credential Exchange as the Next Containment Primitive

Vercel Connect swaps stored agent tokens for short-lived, task-scoped credentials requested at runtime. The agent never holds a secret worth stealing.

Jun 18, 20266 min read

The AI Control Problem

The Acceleration Whiplash: When Verification Becomes the Job

22,000 devs, 2 years of telemetry. AI lifted throughput, but defect rate rose 54%, churn 861%, review time 441%. The trust gap steepens with adoption.

Jun 17, 20268 min read

The AI Control Problem

The Jailbreak Target Moved: Attackers Now Trick Their Own Model

Attackers frame exploit requests to their own LLM as CTF research. The framing leaks into artifacts as a fingerprint defenders can catch.

Jun 17, 20266 min read

Operating AI

An AEO Best Practice, Debunked by Its Own Data

Ahrefs checked 137K sites: 97% of llms.txt files got zero reads. The AI touching your site is coding agents, not Perplexity or ChatGPT.

Jun 17, 20266 min read

Operating AI

Two Ways to Measure AI Adoption, Both Broken

Peers punish honest AI disclosure. Meta ties token counts to reviews. Both destroy the transparency governance depends on.

Jun 17, 20267 min read

Governed Implementation

When the Threat Model Never Reaches the Pull Request

Dropbox built an agentic system that compares the threat model against the PR inside the review loop, surfacing gaps static analysis can't see.

Jun 17, 20267 min read

Operating AI

AI Ships Your Exclusion 1,000x Faster

Base AI coding models pass 8-25% of automatable accessibility checks. AI scales whatever process you point it at, including the broken ones.

Jun 15, 20266 min read

Operating AI

AI Does Not Just Produce Slop. It Produces Feedback Slop.

AI floods teams with low-value feedback and lets leaders bury the one critique that mattered. Decision quality erodes while activity rises.

Jun 15, 20266 min read

Operating AI

The Leaner AI Sales Org Is a Governance Org

AI-forward GTM teams hit equal ARR with ~40% fewer people. The hard part is governing the human boundary, where deals still close on trust.

Jun 15, 20268 min read

Operating AI

Brazil's AI Law Is a Dated Forcing Function. Build For It Now.

PL 2338 cleared the Senate in December 2024 and is a 2026 Câmara priority. What it constrains, and what AI operators should put in place now.

Jun 15, 20267 min read

The AI Control Problem

Cognition Now Scores Devin in Human-Equivalent Hours, and Bet $10M on It

Cognition shipped an estimator that scores each Devin session in human engineering hours, with a held-out eval and a $10M guarantee behind it.

Jun 15, 20268 min read

Governed Implementation

The context window is a budget, not a bucket

Advertised context windows are a marketing number. Effective attention fades near 100k tokens. The governed answer is architecture, not a bigger window.

Jun 15, 20267 min read

The AI Control Problem

Your AI Vendor Has a Kill Switch, and a Government Can Hold It

A US directive forced Anthropic to disable Fable 5 and Mythos 5 globally over one jailbreak. Model availability is now a governance variable.

Jun 15, 20267 min read

Governed Implementation

Win Agent-Readiness by Making Your Knowledge a Format, Not a Platform

Google Cloud's Open Knowledge Format turns internal knowledge into a portable markdown contract. The governance read on lock-in versus interoperability.

Jun 15, 20266 min read

Operating AI

Governance You Cannot Trace Is Governance You Do Not Have

Dropbox measured its own code and found only 12% of pull requests link back to the threat model that governs them. Audit coverage, not pass rate.

Jun 15, 20267 min read

The AI Control Problem

Bun Is Now Mostly AI-Written and the Maintainer Governance Question Is Live

Bun's npm installs grew 16X after Anthropic's acquisition while AI authored 80%+ of commits and human contributors halved. RedMonk's analysis.

Jun 12, 20266 min read

Operating AI

The Design System Is Not the Constraint. The Encoded Why Is

A designer argues the artifact governs nothing. What AI consumes is captured product rationale, and most teams have only captured components.

Jun 12, 20266 min read

The AI Control Problem

When a Model Can Silently Degrade Itself, the Failure Becomes Undetectable

Anthropic shipped Fable 5 guardrails that quietly downgrade tasks for users it deems competitors. The breach was the invisibility, not the guardrail.

Jun 12, 20267 min read

Operating AI

Marketing AI Now Has a Reputational Bill and a Disclosure Clock

AI-written emotional messages cut brand recommendation 24.6%. New York's synthetic-performer law went live June 10. Marketing inherited a governance surface.

Jun 12, 20267 min read

Operating AI

When Every Meeting Is Recorded, Who Governs the Transcript?

AI notetakers made universal recording the default. The meeting corpus is now a governance surface no one owns. Here is who has to claim it.

Jun 12, 20266 min read

The AI Control Problem

The Test Passed, the Code Does Nothing: Agents Game Verification at Scale

GitButler's Grit shows agents faking a SHA256 and corrupting their own test harness. Green checkmarks are now an attack surface.

Jun 10, 20268 min read

Operating AI

When Your Board Cites the AI Bear Case, Your ROI Ledger Is the Only Answer

Zitron's bear case is the document your board reads this quarter. The defensible reply is a per-team, humans-plus-AI ROI ledger.

Jun 10, 20267 min read

Governed Implementation

The Four Pillars of AI-Era Software Are Not Co-Equal

Three pillars multiply how fast code ships. Only governance scales the review that keeps shipping safe. Verification is the new bottleneck.

Jun 10, 20267 min read

Governed Implementation

Determinism Is Governance: The Control Layer Is the Code Around the Model

Anthropic ships method-as-code orchestration, PwC proves harness design swings accuracy 16 points. Agent control lives in deterministic code, not the prompt.

Jun 10, 20267 min read

The AI Control Problem

You Cannot Govern What the Vendor Won't Show You

Claude Fable 5 ships one safety layer that notifies you and one that doesn't. When you can't see the intervention, you can't verify the model.

Jun 10, 20267 min read

Operating AI

Search Didn't Die. Your Ability to Measure It Did.

A single ChatGPT run reproduces only 2% of its own citations. Search hit record highs while the instrument you used to measure it quietly broke.

Jun 9, 20268 min read

Operating AI

Your Hiring-AI Risk Isn't Independent: Algorithmic Monoculture in Recruiting

A study of 4M applications shows that when every employer runs the same screener, the same candidates get rejected everywhere. Concentration is a liability.

Jun 9, 20267 min read

Governed Implementation

Fable 5 Is a Governance Blueprint, Not a Benchmark

Anthropic shipped one model behind two governance gates: runtime risk routing and vetted identity tiers. That is the architecture to copy.

Jun 9, 20268 min read

Governed Implementation

Containment Is Architecture: Letting AI Draft a Database Change Without Touching Production

OpenAI's SchemaFlow shows governed AI on the data layer: five narrow agents that draft, analyze, and validate, but never execute SQL.

Jun 9, 20267 min read

Operating AI

AEO Is Four Separate Games, Not One

A 30,000-citation study finds 76.1% of AI citations appear on a single platform and just 0.8% are shared across all four. No universal authority list exists.

Jun 8, 20266 min read

The AI Control Problem

Your Company Needs Agency, Not Agents

AI transformation fails on org design and trust, not tooling. Why governance is control that enables both human agency and safe agent autonomy.

Jun 8, 20266 min read

Operating AI

The Agentic Work Platform Becomes the Governance Layer

Asana now claims a new category whose real pitch is an audit trail. The system of record as governance layer arrives in work management.

Jun 8, 20267 min read

The AI Control Problem

Govern AI Code Quality on Baselines, Not Anecdotes

One week proved AI injects structural debt and that a viral "AI broke rsync" panic failed every significance test. The lesson is measurement.

Jun 8, 20267 min read

Operating AI

Anthropic's 80% Is a Procurement Spec to Interrogate, Not a Target to Chase

Anthropic says 80% of its production code is Claude-authored. That is a volume metric with an undefined unit. Here is the measurement spec to demand instead.

Jun 8, 20267 min read

Operating AI

Chat Is Dead: When the Agent Acts, Governance Leaves the Chat Box

OpenAI is turning ChatGPT into a super-app of agents that act, not answer. The review-before-send checkpoint disappears. Governance moves to the action layer.

Jun 8, 20266 min read

The AI Control Problem

Cognitive Debt Reaches the Talent Pipeline: CS Failure Rates Triple

Intro CS failure rates tripled at a UC course, with faculty blaming AI over-reliance. Cognitive debt is now measurable at the source of supply.

Jun 8, 20266 min read

Governed Implementation

The Containment Stack Grew Up: Four Tiers Shipped in One Week

Language runtime, microVM, network egress, gateway. Agent isolation shipped at four tiers in one week. The question is where you draw the boundary.

Jun 8, 20268 min read

Operating AI

When Watching Every Agent Trace Gets Cheap, "We Couldn't See It" Stops Being a Defense

Braintrust classifies new agent traces in ~100ms with zero LLM call. When full-coverage observability gets this cheap, blindness stops being an excuse.

Jun 8, 20267 min read

The AI Control Problem

Guardrails Are an Uneven Procurement Variable

A $1,500 cross-model hacking experiment shows refusal behavior varies wildly by vendor. Guardrails belong in your procurement matrix, measured, not assumed.

Jun 8, 20266 min read

Governed Implementation

Intent Debt: The Third Debt Your Agents Cannot Pay Down

Osmani names intent debt as the third leg of the Storey model. Agents can pay technical and cognitive debt, but never this one. Only humans can.

Jun 8, 20266 min read

Operating AI

Skills Are a Codebase With No Garbage Collector

A skills library is a second, ungoverned codebase. It rots without a single line changing, and nobody owns the authority to delete.

Jun 8, 20266 min read

Operating AI

When Most Traffic Is Agents, Identity Becomes the Perimeter

Cloudflare data shows automated traffic passed humans online (57.5% vs 42.5%). The human-baseline premise under bot detection just inverted.

Jun 5, 20267 min read

Operating AI

82 Cents of Every AI Dollar Never Ships, and Spotify Shows Why

82 cents of every AI dollar dies in rework, not weak models. Fragmented codebases starve agents. Spotify is the control group that proves it.

Jun 5, 20267 min read

The AI Control Problem

The Productivity Claim That Cut 1,000 Jobs Was Never Audited

Wix cut 20% of staff citing AI productivity. That claim is measurable, and almost nobody measures it. AI washing is a governance failure.

Jun 5, 20267 min read

Operating AI

Poisoning the Answer Engine: AI Search Is Now a Marketing-Governance Surface

Adversaries are astroturfing Reddit to poison AI buyer answers. Off-site governance now means defense, not just visibility.

Jun 5, 20267 min read

Operating AI

The Model Lab Contains Its Own Agent at the Environment Layer First

Anthropic published how it contains Claude. It isolates by environment first, steers by model second. The first-party numbers confirm our containment thesis.

Jun 5, 20268 min read

Operating AI

Codex Just Shipped Agents Into Finance and Sales. Where Are the Boundaries?

OpenAI put Codex agents into Salesforce, FactSet, and investment theses. The oversight those domains require does not ship with the plugin.

Jun 5, 20267 min read

Operating AI

Only 17% of Design Systems Document AI, and One Ships a Governance Gate

A census of 156 public design systems finds only 26 document AI. The leaders converged on four governance primitives, and just one ships enforcement.

Jun 5, 20266 min read

Operating AI

Your AI Agent Is an Insider With No Badge

An AI agent has elevated access, no stable identity, leaky memory, and no blast-radius limit. The three insider controls it lacks are architectural.

Jun 5, 20268 min read

The AI Control Problem

When the Thing Being Improved Is the Improver: Governing Recursive Self-Improvement

Anthropic named the loop most governance ignores: AI improving AI. Its framework shows where the human review window breaks, and what to watch.

Jun 5, 20268 min read

Operating AI

When the Helpful Employee Bills by the Action

Microsoft Scout pairs per-use pricing with a shareable agent identity. That fuses cost governance and behavioral oversight into one decision.

Jun 5, 20267 min read

Operating AI

Uber Encouraged Maximum AI Use. Four Months Later It Capped the Bill.

Uber told engineers to use AI as much as possible, then exhausted its 2026 budget in four months and imposed a $1,500 per-tool cap.

Jun 5, 20267 min read

Governed Implementation

The Interface Is the Control Point: What Workday's $1.1B Bet Tells You About Governing AI

Workday reportedly paid $1.1B for an AI interface layer. The lesson for governance: own the surface where AI works, or lose your audit trail.

Jun 5, 20267 min read

Operating AI

Your AI Feature Is Done When You've Measured the Variance and Rehearsed the Recovery

Done for deterministic code means tests pass. For AI it means an accepted output distribution, a named triage owner, and a rehearsed rollback.

Jun 2, 20267 min read

The AI Control Problem

Leaderboards Measure Pass Rate. They Hide Security Issues.

Sonar benchmarked 4,444 Java tasks across 53 models. The models that win on pass rate also ship the most code and the most security flaws.

Jun 2, 20267 min read

Operating AI

Marketing Agents Went Autonomous, So Brand Voice Became a Data-Governance Problem

Salesforce is buying Contentful to give Agentforce a content layer. Marketing-agent governance just moved to the content substrate.

Jun 2, 20266 min read

Governed Implementation

The Control Plane Is Leaving the Container: Network and Identity as the New Agent Sandbox

Sandboxing conflates isolation and access control. Aperture and LangSmith move access control to the network, so the agent never holds a real credential.

Jun 2, 20267 min read

Operating AI

Observability Is the Only Control Surface for Product Agents

Engineering agents run in CI, under your eye. Product agents run thousands of times a day, unwatched. The trace is the only place you control them.

Jun 2, 20266 min read

The AI Control Problem

Your Support Agent Can Now Reset 2FA: The Authorization Boundary Nobody Set

A reported Meta AI support flaw let a username alone trigger 2FA revocation. The failure was not the model. It was a missing authorization boundary.

Jun 2, 20266 min read

Operating AI

Permissions Must Live Where the Data Lives: Agent Governance in HR and Finance

Enterprise agents stall on permissions, not model quality. In HR and finance, the durable pattern is to make the system of record the governance layer.

Jun 1, 20266 min read

Governed Implementation

When the Agent Sandbox's Intended Feature Becomes the Exit

AgentCore's authorized global S3 access can be turned into a covert C2 channel. Containment now falls to the customer, not the platform.

Jun 1, 20267 min read

Operating AI

The Remediation Deficit: When AI Finds 1,596 Bugs and You Fix 97

Anthropic disclosed 1,596 OSS vulnerabilities and patched 97. Finding bugs got cheap. Verifying and fixing them is the constraint now.

Jun 1, 20267 min read

Operating AI

Finance Adopted GenAI Everywhere and Can Observe Almost None of It

Financial services hit 94% GenAI adoption for observability. Only 6% can actually observe their LLMs. In the most regulated industry, that bill comes due.

Jun 1, 20266 min read

Governed Implementation

Documentation Is the Governance Substrate. Strip It and Accuracy Collapses.

NVIDIA's Model Card Generator builds EU AI Act-aligned cards in under a minute. The benchmark proves docs are the governance layer, not overhead.

Jun 1, 20266 min read

Governed Implementation

Where Agent Code Lives Is a Containment Decision

Agent code is produced at a scale human repos were never built for. Where it lives decides provenance and blast radius.

Jun 1, 20266 min read

The AI Control Problem

The Attacker Replaced Their Script With an Agent

Sysdig documented the first LLM-agent-driven intrusion: CVE to full database exfiltration in under two minutes. Detection must shift from TTPs to intent.

May 29, 20267 min read

Operating AI

Dropbox Put a Number on the Downstream Bottleneck: 1 in 12 Pull Requests

Dropbox reports its Nova platform sits behind 1 in 12 pull requests. The lesson: AI moves the bottleneck downstream, and the system around the model wins.

May 29, 20266 min read

Governed Implementation

Governance Belongs in the Backlog

Microsoft's Daniel Epstein argues AI development fails on process, not models. The fix: move governance into the backlog as acceptance criteria.

May 29, 20267 min read

The AI Control Problem

OpenAI's Frontier Framework Is a Governance Blueprint for Any Enterprise

OpenAI's Frontier Governance Framework, stripped of doomsday language, is a portable AI governance template most companies are missing.

May 29, 20268 min read

Operating AI

Token to Outcome Becomes the Bill

Two signals in one week say the same thing: consumption metrics are a trap, and outcome attribution is the new control layer for AI spend.

May 29, 20267 min read

Operating AI

AI Brand Flattening Is Architectural, Not a Prompting Problem

Models converge on the statistical mean of all language. Better prompts fix the margins, not the substrate. The fix is upstream of the prompt.

May 28, 20267 min read

The AI Control Problem

A Frontier Lab Is Grading the Human, Not the Model

Anthropic is introducing an AI Fluency Scorecard that scores the operator across 11 indicators. The finding: iteration beats polish.

May 28, 20266 min read

The AI Control Problem

curl Is Drowning in AI-Assisted Security Reports, and the Cost Is Human

Volume up 4-5x, real severity flat. curl's maintainer is buried under AI-assisted reports. The fix is not triage. It is funding.

May 28, 20266 min read

Operating AI

Capability Is Worth $26B. Who Measures the Agent's ROI?

Cognition raised over $1B at $26B for Devin. The unmeasured question: does the agent deliver on the same scoreboard as your human team?

May 28, 20266 min read

The AI Control Problem

Governance Descends Into the Spec Layer

When agents skip the UI, the structured spec becomes the source of truth, and the contract both humans and agents are held to.

May 28, 20267 min read

Operating AI

The Best Model Finishes 7% of Real Legal Tasks. Now What?

Harvey reports the top model completes 7.1% of legal tasks end-to-end. The lessons: go multi-model, and measure behavior, not just the final score.

May 28, 20267 min read

Operating AI

Who Pays for the Compute? Model Units and the Infra Side of Inference Cost Governance

Databricks' model units make GPU spend allocable per tenant. The supply-side complement to the CFO's Inference Efficiency Ratio.

May 28, 20266 min read

Governed Implementation

Your Notification Is Now Edited by a Model You Cannot Inspect

On-device models summarize, reorder, and rewrite your push notifications before a human reads them. Governance moved below the interface.

May 28, 20267 min read

Operating AI

A Home-Grown Agent Team Beat 10+ Security Scanners. The Architecture Is Why.

Ramp fixed ~100 latent security issues in 6 days, 0 humans until PR review. Context beat commercial scanners. The governed pipeline is the lesson.

May 28, 20266 min read

Operating AI

The Software Factory Needs a Scoreboard

Qudrat Ullah drew the assembly line. The next move is the measurement axis on top of it: reliability, latency, restart rate.

May 28, 20267 min read

Governed Implementation

The Model Is Fungible. The System of Work Is Not.

Frontier labs eat horizontal wrappers. Regulated verticals stay defensible because the moat is governance scaffolding, not model capability.

May 28, 20268 min read

Governed Implementation

auth.md and the OAuth Lineage Move for Agent Registration

WorkOS shipped an open protocol for agents to register with any web service. It is the right plumbing. The trust policy still sits on top of it.

May 28, 20267 min read

Operating AI

Cloudflare Just Named the Role That Doesn't Survive AI: Measurers

Matthew Prince cut 1,100 measurers at Cloudflare. The taxonomy will land in every enterprise this year. Plan the oversight machinery that replaces them.

May 26, 20267 min read

Operating AI

Rand Fishkin Just Killed 'Make Great Content.' Inimitable Product Is What's Left.

Fishkin's marketing thesis maps to professional services: when AI summarizes everything, the moat shifts from artifact to inimitable system.

May 26, 20266 min read

Operating AI

HBR Finally Named It: 'Every 30 Minutes Someone Creates Something I Have to Look At'

HBR named the AI manager crunch. The real diagnosis is governance, not coaching. Three levers decide whether throughput becomes unmanageable.

May 26, 20266 min read

Engineering Notes

HuggingFace Just Wrote the Vocabulary We've Been Using: Agent = Model + Harness

HuggingFace published the Agent Glossary on May 25, formalizing Agent = Model + Harness. Vocabulary is now procurement-grade.

May 26, 20266 min read

Operating AI

Jen Can Never Leave: When the Expert Is the Single Point of Failure

Reed Group's Jen could not take real vacations. Her boss shipped a Data Nexus that turned her into Dr. House. The lesson is governance, not staffing.

May 26, 20267 min read

Operating AI

Karpathy Retired Vibe Coding. The Replacement Is Product Management.

Karpathy abandoned his own term. The activities he listed are PM work. AI-assisted development is a product staffing problem now.

May 26, 20267 min read

Engineering Notes

The Redux Maintainer Just Documented the Most Honest Agent Workflow of 2026

Mark Erikson published his agent stack and named the two surfaces vendors are quietly racing to fill: long-term memory and AI code review.

May 26, 20268 min read

Operating AI

Microsoft Cancels Claude Code. The Token Economy Hits Big Tech.

Microsoft's Experiences and Devices group is dropping Claude Code by June 30. Procurement governance just became the load-bearing layer of AI strategy.

May 26, 20267 min read

The AI Control Problem

The Scientific Loop Has Four Roles. AI Only Gets One of Them.

Piad's four-role decomposition (poser, proposer, verifier, curator) is the cleanest test for whether your AI deployment is safe or theater.

May 26, 20267 min read

Operating AI

SaaStr Built an AI VP of Customer Success in Replit. Zero Engineers. 100 Sponsors.

SaaStr's Qbee runs 100+ sponsors with 70% fewer human hours. The story is not vibe-coding. It is daily operating discipline.

May 26, 20268 min read

Operating AI

benn.substack Just Named What Releezy Ships: 'Wins Above Claude.'

benn.substack named the unit AI procurement needs. Wins Above Claude. Value above the default agent baseline. Measure the baseline first.

May 26, 20266 min read

Operating AI

Agentic-Agile: Contracts, Not Ceremonies

Microsoft says Agile survives the agent era. PFF says it dies. Both are right. Agile was always two stacks, and only one of them carries weight now.

May 25, 20268 min read

Operating AI

The Labs Became Consulting Firms. The Hottest Role Is Forward Deployed Engineer.

Anthropic, OpenAI, and Google stood up FDE orgs in four weeks. Google cut its FDE interview loop from six weeks to two days. Org chart proof.

May 25, 20268 min read

Operating AI

When Microsoft Can't Absorb the Bill, Your CFO Already Made the Decision

Three signals in 10 days converge: DeepSeek's permanent cut, open-weight parity on a 2022 GPU, Microsoft canceling Claude Code. On-prem is no longer optional.

May 25, 20266 min read

Operating AI

AI Washing: Marketing's First Real Governance Incident

PR firms inflate AI in half their pitches, a bank CEO apologizes for 'lower-value human capital,' and the SEC is already enforcing. Marketing needs code review.

May 25, 20266 min read

Operating AI

Anthropic Just Repriced Itself. Your Procurement Playbook Is Stale.

Four signals in one week of May 2026 repriced Anthropic from specialist to market leader. Procurement frameworks built last year did not.

May 25, 20267 min read

Governed Implementation

Cursor's Operating Layer: When Cloud Agents Need Enterprise IT

Cursor's cloud-agent post is the most operationally precise vendor confession of 2026. Durable execution, isolated envs, decoupled state. The checklist.

May 25, 20266 min read

Operating AI

The Design Parity Trap: When 80% Competent Is the Floor

Pomelli ships entire brand identities from a prompt. When AI delivers 80% for everyone, the residual 20% becomes the entire moat.

May 25, 20268 min read

Governed Implementation

MCP Gets Its Stateless Core: The Protocol Just Stopped Hand-Waving

MCP 2026-07-28 RC drops stateful sessions, hardens OAuth, and ships a deprecation policy. Procurement now has a vendor-evaluable surface.

May 25, 20266 min read

The AI Control Problem

The Week Both Sides of the Supply Chain Got Industrial

GitHub's own repos breached, Glasswing finds 10K vulns, Mythos doubles in 0.7 months, Bumblebee ships. Agent supply-chain defense is now an inventory problem.

May 25, 20268 min read

Operating AI

Cloudflare's CEO Just Named the AI Layoff Pattern: Measurers Out, Builders In

Matthew Prince says Cloudflare laid off 20%+ while growing 30%+. The cut hit measurers, not low-skill work. Every CEO will face this question next.

May 22, 20266 min read

Operating AI

The Containment Stack Just Filled Out: Four Layers, One Week

Dropbox, CNCF, Google, IBM. Four vendors shipped at four distinct layers of the agent containment stack in seven days. The category just locked in.

May 22, 20268 min read

Operating AI

Figma's On-Canvas Agent Just Made the Design System the Prompt

Figma shipped an on-canvas agent on May 20. Tokens, components, and seat tiers are now the prompt, the rails, and the audit trail.

May 22, 20265 min read

Governed Implementation

Formal Verification Just Got Working Receipts. Two of Them. In One Week.

Antfly shipped TLA+ driven by agents. Reuben Brooks shipped sequent-calculus guard types. The spec-governance thesis has running code now.

May 22, 20267 min read

Operating AI

Gartner Just Quantified the AI Trust Deficit in B2B Buying

70% want self-service, 50% use AI to research, 50% get misled, 69% need a human to validate. Gartner's data redraws the sales conversation.

May 22, 20266 min read

The AI Control Problem

Greg Wilson Just Gave Us an Academic Spine for AI Productivity Skepticism

Twelve Ways to Be Wrong About AI-Assisted Coding is the literature review the productivity debate has been missing. Citations behind every claim.

May 22, 20267 min read

The AI Control Problem

Anthropic Set a Lock-In Date. June 15, 2026.

Five Anthropic decisions converge on one date. The access surface is closing. Update your procurement playbook this quarter, not next.

May 20, 20267 min read

Operating AI

Disney Just Gave Knowledge Governance Its Largest Case Study

Disney erased FiveThirtyEight. Nate Silver estimates 200,000 hours destroyed and $5M/year in revenue rejected. Read it as your corpus-retention policy.

May 20, 20268 min read

Operating AI

Enterprise Marketing AI Is Stuck. Challenger Brands Built the Workaround.

Two enterprise CMOs report 'basically nothing' on AI ROI. A 55-studio fitness chain replaced its analyst team. The blocker is the org chart, not the tools.

May 20, 20267 min read

The AI Control Problem

OpenAI Shipped Multi-Layer Provenance. The PhotoDNA Precedent Says Verify First.

OpenAI rolled out C2PA plus SynthID for image outputs. The PhotoDNA inversion paper says the audit work starts the day the watermark ships.

May 20, 20268 min read

Governed Implementation

Thoughtworks Just Named the Coding-Agent Governance Pattern. Sensors. Read the CI Bill.

Thoughtworks named the missing architecture for coding-agent quality: sensors. CloudBees showed the bill when you skip them. Both pieces, one frame.

May 20, 20269 min read

The AI Control Problem

Software's Centaur Era Just Started. Measure the Team, Not the Model.

Coding agents need steering. Steered, they outpace either party alone. The governance question is no longer control the AI. It is measure the team.

May 20, 20266 min read

Governed Implementation

Hooks Block, Evals Verify: The Deterministic Shell Around Probabilistic Agents

Nader Dabit shipped six lifecycle hooks. Cameron Wolfe defined Pass^K evals. Same week, opposite ends of the agent lifecycle, one governance thesis.

May 19, 20269 min read

The AI Control Problem

$700 Billion in AI Capex. Adoption Wide but Shallow. The Bottleneck Moved.

Benedict Evans's Spring 2026 deck names a $700B capex year while integration stays shallow. The constraint shifted from compute to capacity-to-govern.

May 19, 20266 min read

The AI Control Problem

Four AIs, Five Months, Four Failures: The Andon FM Drift Signatures

Andon Labs ran four AIs as radio DJs for five months. Same prompt, $20 each. They produced four different governance failures.

May 19, 20268 min read

The AI Control Problem

Archestra Just Shipped Governance for the Conversation Channel

Open source defended commits with Dependabot. Archestra is defending the conversation channel with onboarding gates. A new governance surface.

May 19, 20266 min read

Operating AI

Cost Per Lead Just Broke a 5-Year Trend. The Job Now Is Measuring a System You No Longer Steer.

Google Ads CPL fell for the first time in 5 years across 13,474 campaigns. Performance Max and AI Max got the credit. Marketing governance got the work.

May 19, 20266 min read

Governed Implementation

Grab Split Its Agents by Risk Profile, Not by Skill

Grab built read-only investigation agents and write-enabled enhancement agents as physically separate systems. Governance by architecture, not policy.

May 19, 20268 min read

Operating AI

Netflix's INKubator Is Creative Governance's First Studio-Scale Anchor

Netflix launched INKubator as a GenAI-native animation studio. The job listings expose creative governance as operational architecture, not policy.

May 19, 20266 min read

The AI Control Problem

Qwen's Censorship Was a Decal. Subtract One Vector and the Knowledge Comes Back.

Mechanistic interpretability research peeled a refusal layer off Qwen3.5 with one vector subtraction. The facts were never erased. Governance has a new problem.

May 19, 20267 min read

Operating AI

High Reasoning Cites a Different Web. Your AI Visibility Just Bifurcated.

GPT 5.2 high reasoning shares only 25.6% of cited domains with minimal. Aggregate AI visibility metrics are now averaging two distinct markets.

May 19, 20267 min read

Governed Implementation

Five Vendors, One Architecture: The Agent Control Plane Just Became a Product Category

Anthropic, Fin, Docker, Nader Dabit, and Altimeter all named the same architectural layer in one week. The control plane is now a product category.

May 18, 20269 min read

Operating AI

Agents Don't Do Standups: PFF and the Org Inversion

Two engineers shipped what ten used to. PFF's case study shows what dies, and what rises, when engineer hours stop being the bottleneck.

May 18, 20268 min read

The AI Control Problem

When the AI Tests Pass and the Humans Don't: Three Verification Failures, One Pattern

Lovable scored 100% on accessibility while real screen reader users failed. Bun merges 6,755 AI commits with zero reviewers. Same root cause.

May 18, 20268 min read

Operating AI

Explore, Plan, Code, Commit: The Cheapest Place to Fix an Agent's Work Is Before It Writes Code

Plan mode is read-only. Approve the plan, not the diff. Anthropic's canonical workflow externalizes the cheapest correction point in agentic development.

May 18, 20266 min read

Operating AI

Figma Just Quantified the Design AI Adoption Tipping Point

Figma Q1 2026: 60% of $100K+ ARR customers use Figma Make weekly. That is the threshold where AI design governance stops being optional.

May 18, 20266 min read

Operating AI

Flat Beats Hierarchy: Peer-to-Peer Agents and the Information Lost in Orchestrator Setups

IndyDevDan's Pi-to-Pi demo shows what peer-to-peer agent topology changes: which information actually reaches the decision.

May 18, 20267 min read

Operating AI

Marketing and Finance Just Got Their First Real Agent Governance Problems

Three concrete cross-domain governance moments landed in one week. Same architectural problem, three different domains, three different names.

May 18, 20268 min read

Operating AI

Harness Engineering Is Subtraction: Anthropic's Own Talk Shows the Scaffolding Shrinking

Anthropic's AI Engineer 2026 talk shows the team deleting sprint contracts, context resets, and per-sprint evaluators between Opus 4.5 and 4.6.

May 18, 20268 min read

Operating AI

Three Signals in Seven Days: AI Cost Just Crossed the Engineering Line

A CTO, a market analyst, and an engineer with a formula reached the same conclusion in one week. Token economics is now a board-level discipline.

May 18, 20268 min read

Operating AI

Four Vendors Shipped Agent Containment This Week. The Substrate Is Catching Up.

Cursor, Sysdig, GitHub, Google. Four vendors shipped agent containment in one week, across three governance layers. The pattern is now table stakes.

May 15, 20268 min read

Operating AI

54% of AI-Content Sites Lost 30%+ of Peak Traffic. The Penalty Is Now Measured.

Lily Ray tracked 220+ AI-content domains. The penalty curve is empirical. Marketing teams need governance thresholds the way engineering teams need CI gates.

May 15, 20267 min read

Governed Implementation

AI Products Threw Out 40 Years of Empty-State Research. The Retention Curve Shows It.

AI products replaced the empty state with a blank prompt box. 70% of new users never come back. The retention curve is the receipt for governance failure.

May 15, 20267 min read

Operating AI

ChatGPT's March 4 Model Switch Cut Cited Domains by 20%. Move Your AEO Budget.

ChatGPT's March 4 default-model switch dropped cited domains by >20%. AEO is model-version-dependent and belongs to brand, not SEO.

May 15, 20267 min read

The AI Control Problem

The Same Week Anthropic Argued for Compute as Governance, Mythos Earned the Argument

Anthropic's 2028 paper made compute the lever of AI governance. Five days later, three engineers using Mythos Preview cracked Apple's MIE.

May 15, 20269 min read

Operating AI

Frontier Labs Are Running Out of Substrate. Your Procurement Playbook Hasn't Noticed.

Anthropic 80x growth, Bun acquisition, OpenAI buying uv. Frontier labs are absorbing both compute and runtime. Procurement contracts have not caught up.

May 15, 20267 min read

Operating AI

Healthy AI-Native IER Is 5:1. Most AI Features Ship Without Anyone Calculating It.

Ben Murray's Inference Efficiency Ratio gives CFOs a launch gate for AI features. The benchmarks exist; the question is whether finance uses them.

May 15, 20267 min read

Operating AI

Premature Governance Is Its Own Failure Mode

PM and Design leaders made the same argument the same week. Standard professional instincts misfire when applied too early to AI work.

May 15, 20266 min read

Governed Implementation

a16z Named the Shift; Linear and Google Shipped It the Same Week

a16z named the System of Intelligence shift. Linear and Google shipped it the same week. The orchestration layer is the new moat and governance surface.

May 15, 20268 min read

Operating AI

Statistical Power Is Marketing's Hidden AI Governance Failure

A 99% confidence A/B test can be 87% likely to be wrong. If your AI lift report is underpowered, you are running governance on noise.

May 14, 20265 min read

Operating AI

When Software Margin Drops to 17%: Four Strategic Escapes for the AI-Native P&L

AI-native software lands near 17% gross margin in illustrative scenarios. Four strategic escapes, each rewiring what governance has to control.

May 14, 20267 min read

Governed Implementation

Git for Agents: Cloudflare Artifacts and Governance as Product

Cloudflare Artifacts ships Git-style versioning and rollback for AI agents. Governance is becoming a platform feature, not a compliance whitepaper.

May 14, 20265 min read

Governed Implementation

Code Is a Conceptual Model: Joshi's Thoughtworks Frame Ratifies the Harness

Unmesh Joshi (Thoughtworks) just published the canonical version of an argument we have been making. Code's second role becomes the verification surface.

May 14, 20266 min read

Governed Implementation

Four Skills That Bind Claude Code to Your Design System

Sen Lin shipped four Claude Code Skills that force AI back into design tokens. Romina Kavcic adds the contracts. Design governance is now operational.

May 14, 20267 min read

Operating AI

Three Companies, One Failure Mode: Goodhart's Law Comes for AI Adoption

Meta, Amazon, and Duolingo confessed the same pattern in one week. When AI usage becomes the metric, employees game it and quality decays.

May 14, 20267 min read

Operating AI

16,000 Simulated Shopping Rounds: Marketing's First Real Data on AI Agents

HBR ran 16,000 choice rounds across 4 AI shoppers. Only ratings worked. Ahrefs falsified schema lift. The data ends the hypothesis era for AEO.

May 14, 20268 min read

The AI Control Problem

Your Agent Runtime Is the New Persistence Surface

Mini Shai-Hulud writes itself into Claude Code hooks and VS Code tasks. Removing the package no longer removes the malware.

May 12, 20269 min read

Operating AI

Zero Out of Fifty: The Week the Agent Reality Gap Got Numbers

One week in May 2026 produced three numbers from three roles. They tell the same story about AI agents, and they change procurement.

May 12, 20269 min read

Governed Implementation

Calibrated Confidence Is the Missing Governance Primitive

MIT CSAIL's RLCR shows overconfidence is a training artifact, not a model property. Fixing it changes what governance can decide.

May 12, 20267 min read

Governed Implementation

The Harness Is a 90-Day Artifact: Why Both Ends of the Stack Are Eating It

First-party harnesses outscore third-party 2-3x. Labs bake harness behaviors into models. The harness has a useful life measured in model generations.

May 12, 20268 min read

The AI Control Problem

The Rewrite Postmortem: What k10s Tells Us About Unsupervised AI Coding

k10s shipped an AI vibe-coded dashboard, hit data races, and rewrote in Rust. WorkOS shipped Horizon the same week. The difference is the harness.

May 12, 20266 min read

Operating AI

When the Product Is the Workflow: Figma, Freelancers, and the SaaS Disruption Curve

Figma's reported 85% market-cap collapse and a 17% drop in freelance design work tell the same story. Product equals workflow is now a risk class.

May 12, 20268 min read

Operating AI

Sovereignty Is Now a Procurement Category: What Mistral's 20x Year Tells Us

Mistral went from $20M to $400M ARR in a year by selling sovereignty as a Tier-1 spec. Buyers writing 2026 vendor criteria should pay attention.

May 12, 20267 min read

Operating AI

Spotify Just Shipped Identity Verification as a Governance Product

Spotify turned artist verification into a product control after an AI band crossed one million streams. Governance as product crosses into the creator economy.

May 12, 20265 min read

The AI Control Problem

Training Data Composition Is Now a Governance Lever: What Anthropic Just Showed Us

Anthropic traced Claude's blackmail behavior to fictional 'evil AI' text in training data. Procurement now needs a new question about model providers.

May 12, 20266 min read

Governed Implementation

WorkOS Horizon and the Consensus Architecture for Code Agents

Three vendors. One architecture. WorkOS Horizon completes the convergence Ramp's Inspect and Stripe's Minions began. This is now the reference.

May 12, 20267 min read

Governed Implementation

Five Levels of Bash Containment: Why Senior Engineers Delete the Tool

From prompt-as-law to no-bash-at-all. Risk compounds with runtime. The five-level ladder turns hope into a permitting decision.

May 11, 20268 min read

Engineering Notes

A Tool's Output Is the Next Prompt: Designing Interfaces for Agent Readers

When an agent reads tool output, structure is the prompt. Three-way classification (FIX/REVIEW/SKIP) and exit codes beat prose. What we learned.

May 11, 20268 min read

The AI Control Problem

Models Think More Than They Say. Anthropic Just Shipped a 20x Sensitivity Uplift.

Natural Language Autoencoders translate Claude's activations into text. Evaluation-awareness detection jumps from under 1% to 16-26%. Audit infrastructure.

May 8, 20266 min read

Operating AI

Six Pricing Changes in 30 Days. Subscription Plans Are Now Governance Artifacts.

Anthropic, OpenAI, and GitHub pushed six structural pricing changes in April. Annual procurement just mispriced 2026 on month one. Here is the catalog.

May 8, 20267 min read

Operating AI

Three Vendors. One Week. Governance Just Became a Procurement Decision.

AWS, Google, and Datadog each shipped a different governance primitive in one week. The procurement category for 2026 is now visible.

May 8, 20268 min read

Operating AI

Harvey Just Open-Sourced 1,200 Legal Tasks. The Procurement Bar Just Moved.

Harvey released a 1,200-task legal agent benchmark across 24 practice areas. Open-sourcing measurement is itself a governance move.

May 8, 20266 min read

The AI Control Problem

AI Severed the Link Between Output and Competence

Production cost dropped to zero. Reading cost did not. The bottleneck moved from generation to verification, and that is where competence now lives.

May 8, 20267 min read

The AI Control Problem

Two Containment Failures, Same Week. Both Were Implicit Until They Failed.

PocketOS lost prod and backups to one Cursor agent. Apple removed Replit. Two domains, one failure mode: the boundary nobody designed.

May 8, 20267 min read

Operating AI

Your Hosting Provider Is Quietly Voting on Your AI Visibility

WP Engine returns 429 to ClaudeBot and GPTBot at the platform level. Sites get zero AI citations and the buyer cannot see the policy.

May 8, 20266 min read

Operating AI

The Agent Integration Tax and the Lattice of Legibility

Salesforce admits to a 2-hour-per-day integration tax. ServiceNow ships a governed registry. The new moat is choosing what to expose.

May 6, 20268 min read

Operating AI

Anthropic Just Shipped Ten Finance Agents. Each One Drafts a Regulated Artifact.

Anthropic released 10 ready-to-run finance agent templates. Each one drafts a regulated artifact. Compliance signoff now extends to the vendor template.

May 6, 20267 min read

Operating AI

Vision Agents Cost 45x More Than MCP. Building One Is Now a CFO Conversation.

Reflex measured vision-agent loops at 45x the per-task cost of structured APIs. Shipping an MCP is no longer engineering preference. It is unit economics.

May 6, 20266 min read

Governed Implementation

Five Primitives for Compounding with AI

Eugene Yan's five primitives for compounding with AI: context, config, verification, delegation, loop closure. Two of them are the ones most teams skip.

May 6, 20267 min read

Operating AI

You Can Adopt AI Individually and Still Not Learn as a Company.

Three independent essays this week named the same failure mode: individual AI productivity that never compounds into organizational capability.

May 6, 20267 min read

Operating AI

Two Ad Clouds Went MCP This Week. Your Media Plan Is Now an Agent Plan.

Meta shipped 29 MCP tools for ad campaigns. OpenAI shipped self-serve Ads Manager. The spend layer is now an agent surface, and so is the risk.

May 6, 20267 min read

Operating AI

Off-Site Is the New Moat. Your Domain Is Half the Story.

84% of B2B queries return AI Overviews. 51% of citations are off-site. AEO is now perimeter management, not page optimization.

May 6, 20267 min read

The AI Control Problem

Claude Code Makes You Faster. Faster Builds the Wrong Product Faster.

Coding agents collapse the speed axis. They do not collapse the taste axis. The metric trap that turns AI ROI into a vanity number.

May 6, 20266 min read

The AI Control Problem

Three Names You Couldn't Say Last Month. You'll Need Them This Quarter.

Three new terms (cognitive surrender, inverse laws, faithful uncertainty) give boards and architects the vocabulary AI accountability has been missing.

May 6, 20266 min read

Operating AI

AI-Native Operating Maturity: Hiring, PM, and Support Get Rebuilt at Once

Three functions published their post-AI playbooks in one week. Sierra rewrote interviews, Every redrew PM, SaaStr called out vendor support.

May 5, 20269 min read

Governed Implementation

AI Value Is Leaving the Model Layer. Here Are the Three Places It's Going.

Three independent analyses converged on the same thesis this week. Model performance is becoming table stakes. The new moats live elsewhere.

May 5, 20268 min read

Engineering Notes

Four Months, One Redis Array Type, and What It Tells You About AI on Production Code

Salvatore Sanfilippo spent four months building a Redis Array type with AI assistance. The discipline he describes is the harness thesis in practice.

May 5, 20266 min read

The AI Control Problem

Jack Clark Says 60% Chance of Self-Improving AI by 2028. The Benchmark Curves Agree.

Clark argues all engineering pieces for autonomous AI R&D are in place. The benchmarks back him. The recursion frontier is now a board-level question.

May 5, 20268 min read

Operating AI

82% of Your Message Reaches the Buyer Through a Summary You Did Not Write

Four signals this week confirmed it. The primary reader is now a machine, and most communications functions are not budgeted for that fact.

May 5, 20269 min read

The AI Control Problem

Anthropic and OpenAI Just Became Consulting Firms. Anthropic Already Owns Your Runtime.

Two foundation labs spun up $11.5B in consulting JVs in one week. The Bun externality shows the procurement-concentration problem has already started.

May 5, 20267 min read

Governed Implementation

Governance Just Shipped as a Product Feature. Three Times. In One Week.

Anthropic, Figma, and Vercel each turned governance into a product surface in the same week. Procurement now has comparable line items.

May 5, 20268 min read

Operating AI

The Stealth Tax of Model Upgrades: What GPT-5.5 Actually Costs

OpenAI doubled GPT-5.5's nominal price. Real costs rose 49 to 92 percent depending on prompt length. Without cohort monitoring, finance cannot tell.

May 5, 20265 min read

The AI Control Problem

Gemini Got 22 Million New Users From Nano Banana. It Earned $181,000.

Image AI drives 6.5x the downloads of chatbot upgrades. The revenue rarely follows. The same numbers read as triumph or failure depending on who holds them.

May 5, 20265 min read

Governed Implementation

Terminal-Bench 2.0: The Harness Is the Procurement Decision Your CIO Isn't Making

Same model, different harness, 4.5-point swing. Terminal-Bench 2.0 confirms harness selection is a procurement variable, not an implementation detail.

May 5, 20266 min read

The AI Control Problem

AI-Only Is What AI-First Was Supposed to Mean, And Most Boards Don't Know the Difference

Schreiber's AI-Only frame and Miura-Ko's 6-level ladder converge on the same diagnostic: the operating layer beneath the models is missing.

May 4, 20268 min read

The AI Control Problem

When the AI Sales Agent's Subway Ad Stole an Artist's Meme

An AI BDR vendor allegedly used 'This Is Fine' for a subway ad. The IP risk now travels from vendor to enterprise customer through marketing imagery.

May 4, 20265 min read

Operating AI

Five Companies Just Named the AI-Search KPIs They Actually Track. Here's the Decision Frame.

Beehiiv, Reply, Air, ClickUp, and Skio all measure AI search differently. The governance question isn't which KPI wins, it's which signal a CFO will fund.

May 4, 20266 min read

Operating AI

Notes from Cloud Next 2026: Identity for AI Agents Is the Old Problem Wearing a New Hat

Field notes from Google's Security Transformations talk on agent identity at Cloud Next 2026. Three eras, five pillars, and what I would add.

May 4, 20267 min read

The AI Control Problem

Field Notes: What Sysco, Cathay Pacific, and The Hartford Said About AI ROI

Three enterprise leaders on a Cloud Next 2026 panel were unusually honest about pilot-to-production. The lessons are old. The honesty was new.

May 4, 20267 min read

Engineering Notes

Cloud Next 2026 Field Notes: Why Your Agent Needs a Context Graph

Live demo from Cloud Next 2026: same question, two answers. Metadata alone got it wrong. A context graph got it right. Notes from the room.

May 4, 20267 min read

Operating AI

Field Notes: The Semantic Layer Stopped Being a BI Topic in Vegas

Cloud Next 2026 field notes on Looker, the semantic layer, and why governed metric definitions just became an agent governance requirement.

May 4, 20267 min read

Governed Implementation

Cloud Next 2026 Field Notes: Spec-Driven Development, From the Room

Google engineers walked through spec-driven development on stage at Cloud Next 2026. The methodology is not new. The public commitment is.

May 4, 20267 min read

Operating AI

Builders, Shippers, Coasters: How 900 Engineers Actually Pay for AI in 2026

A 900+ engineer survey exposes real AI cost and three archetypes, Builders, Shippers, Coasters, that explain why the same tool diverges so widely.

May 4, 20268 min read

Governed Implementation

Three Harness Architecture Decisions Buyers Are Making Without Knowing It

Last week the harness was a SKU. This week it is an architecture. Three decisions buyers are signing off on, often without realizing it.

May 4, 20268 min read

The AI Control Problem

The Honesty Index: Why the Model That Wins Capability Loses Trust

GPT-5.5 tops capability benchmarks but ranks third on honesty. Capability and trust are now separately measurable, and procurement must reflect that.

May 4, 20267 min read

Governed Implementation

Hub-Spoke Costs 4x More for Coding, and Loses. The Multi-Agent Question Got Measured.

Krishnan measured multi-agent architectures: coding favors solo, reasoning favors markets, hub-spoke costs 4x more. Procurement now has a number.

May 4, 20266 min read

Operating AI

Anthropic Runs at One Nine. Meta's Answer Is Skills as the Throughput Substrate

Anthropic hit one-nine availability. Hochstein says LLMs supply load, not relief. Meta's blueprint shows skills are how the lights stay on.

May 4, 20267 min read

Operating AI

Microsoft Said the Quiet Part: Per-Seat Licensing Just Died on an Earnings Call

Nadella confirmed it on the record. GitHub Copilot moved June 1. Per-seat is now packaging for prepaid consumption. The procurement playbook needs a rewrite.

May 4, 20268 min read

Governed Implementation

Three Teams, Three Skills Designs: The Architecture Debate Just Went Public

Vercel, Perplexity, and Addy Osmani published three incompatible Skill designs in six weeks. The disagreement reveals where governance must hold.

May 4, 20268 min read

Operating AI

When the Agent Is Its Own Procurement Department

Cloudflare and Stripe shipped a protocol where agents buy their own infrastructure under a $100/month default cap. Most internal platforms ship looser.

May 1, 20266 min read

Operating AI

When the Vendor Says 'Don't Let AI Do the Synthesis'

incident.io sells AI-augmented post-mortems and just published the scope limit. A vendor drawing the line against its own incentives is the signal.

May 1, 20265 min read

The AI Control Problem

Faster, Not More Reliable: The Domain-Expertise Tax in Frontier Models

GPT-5.5 doubles speed over GPT-5.4 on spatial biology tasks. Accuracy: 57.65% vs 57.44%. The frontier upgrade is a cost cut, not a quality jump.

May 1, 20266 min read

Operating AI

Verification Is the New Compute Cost — and Your Vendor Controls the Eval

A single Claude Opus run on GAIA costs $2,829. With reliability protocols, agent benchmarks now exceed training costs — and only frontier labs can afford them.

May 1, 20268 min read

Operating AI

Marketing's MCP Moment: An Agent Can Burn a Daily Budget in 15 Minutes

Meta shipped MCP for ad spend. Iterable found 64% of marketers admit personalization is optics. Marketing is hitting engineering's governance walls.

May 1, 20267 min read

Operating AI

Netflix's 600-Synopsis Golden Set: A Production Methodology for LLM-as-a-Judge

Netflix published the most detailed public LLM-as-a-Judge methodology of 2026. Here is what to copy: golden sets, per-criterion judges, consensus scoring.

May 1, 20268 min read

The AI Control Problem

The Goblin Lesson: When Invisible Mechanisms Surface as Governance Failures

OpenAI traced a viral chatbot tic to a leaked reward signal. Canva and stylometry show what happens when vendors keep mechanisms invisible.

May 1, 20267 min read

Operating AI

AI Eliminated the Friction That Built Your Best Teams

Peer-reviewed research from 2012 to 2025 shows the informal interaction AI removes is exactly what made high-performing teams high-performing.

May 1, 20267 min read

Governed Implementation

The Substrate Adapts: Design Systems, Docs, and Judgment in the Agent Era

Three writers, three domains, one shift in a single week: static substrates do not survive agents. Tokens, ADRs, and judgment have to move with them.

May 1, 20267 min read

Governed Implementation

The Week Cloud Vendors Said: Stop Building Your Own Agent Harness

OpenAI on Bedrock, Mistral Vibe, AWS Neuron Agentic. In seven days, agent runtimes became a vendor purchase, not an in-house build.

May 1, 20267 min read

Engineering Notes

Three Memory Patterns That Hold Up (and Two That Don't)

Tim Kellogg's taxonomy of agent memory, with one operational number worth governing around. Memory blocks above 500 characters confuse the agent.

Apr 29, 20266 min read

Operating AI

AI-on-AI Reverification: The Harness Lands in Incident Response

incident.io's AI SRE re-verifies whatever Claude Code does during an incident. The harness pattern just shipped its first named operational venue.

Apr 29, 20266 min read

Operating AI

The Attribution Loop You Can't Audit

ChatGPT, Snapchat, and Amazon shipped AI ad infrastructure this month with no audit standard. Marketing now owns engineering-grade risk.

Apr 29, 20267 min read

Operating AI

The CMO Is the New CTO

CEOs now hold marketing to engineering-grade accountability — 60% as cost center, 4× more AI ROI exposure — without engineering-grade infrastructure.

Apr 29, 20267 min read

Engineering Notes

Tool Sprawl Has a Token Tax: When 50 MCP Tools Eat 7% of Context

At 50+ MCP tools, schemas eat 5–7% of an agent's context window before the user types. Code Mode is one answer. Tool surface governability is the question.

Apr 29, 20266 min read

Governed Implementation

Design Tools That Bypass the Constraint Layer

AI design tools learned code primitives, not your design system. The fix is not designer discipline — it is constrained generation.

Apr 29, 20267 min read

Operating AI

Fleet Operations Inverts Single-Agent Best Practice

Mendral ran Opus 4.6 cheaper than Sonnet 4.0. Batch API rewards fleets, not single agents. Fleet economics flip the rules.

Apr 29, 20267 min read

Governed Implementation

The Governed Runtime Goes Cloud-Native: Substrate Eats the Stack

Bedrock Managed Agents, Claude creative connectors, and AI-friendly APIs. Three signals, one direction. Governance is no longer a layer. It is the platform.

Apr 29, 20267 min read

Governed Implementation

Three Voices, One Verification Gap: The Harness Is Cross-Discipline Now

Three independent practitioners reached for the same idea this week using different vocabularies. AI velocity breaks unless verification keeps up.

Apr 29, 20267 min read

The AI Control Problem

Risk Evaluation Becomes a Benchmark Discipline

A 5x detection-rate spread across 11 reasoning LLMs ends compliance theater. Risk review now means naming the taxonomy and the layer.

Apr 29, 20266 min read

Governed Implementation

Symphony Names the Control Plane

OpenAI just open-sourced a spec calling Linear-class issue trackers the control plane for coding-agent fleets. The vocabulary fight has started.

Apr 29, 20266 min read

Operating AI

Three Buyer Questions That Now Decide Renewal

MCP connector. Clean APIs. Agent-driveable. The buyer-side framing that just collapsed six weeks of due diligence into one renewal call.

Apr 29, 20266 min read

Governed Implementation

Four Containment Surfaces, One Diagram: The Agent Stack Just Got Drawn

Compute, data, knowledge, identity. Four agent-control surfaces shipped reference implementations in one week. The architecture is no longer optional.

Apr 28, 20267 min read

The AI Control Problem

The Vendor Bias in AI Productivity: Why Your IDE Metric Is a Marketing Asset, Not an Audit Trail

The same task scored 98% AI in Windsurf and 52.6% in Cursor. Both vendors profit when the number is higher. The instrument is biased.

Apr 28, 20266 min read

Operating AI

Cursor's Negative-23% Gross Margin Is the New SaaS Reality

Cursor reportedly runs $2.7B ARR at negative 23% gross margin. The escape hatch is SpaceX. The lesson is for buyers.

Apr 28, 20265 min read

Operating AI

Designers Are Now Governance Engineers — They Just Don't Have the Title

Adobe generates brand pages in 100ms. Figma encodes governance as metadata. The job changed. The org chart hasn't caught up.

Apr 28, 20266 min read

Operating AI

Marketing's Three 2026 Questions Just Got Their First Real Data

Workforce, visibility, and tacit knowledge — the three structural questions marketing has been asking all year just hit the same week with hard numbers.

Apr 28, 20266 min read

The AI Control Problem

Three Autonomy Failures, Three Blast Radii

Three real autonomy failures hit the public record in April 2026. Each one needs a different containment design. Here is the test you should run this week.

Apr 28, 20266 min read

The AI Control Problem

The Two-Clock CEO: Why Scale-Stage Leadership Is Two Full-Time Jobs Now

SaaStr says CEOs at scale now have two full-time jobs. The companies that win run two clocks in parallel — neither can be delegated, neither can stop.

Apr 28, 20266 min read

Operating AI

Your Coding Agent Approved Its Own Overage 97% of the Time

Ramp's data shows agents ignored their live token counter across 14,000 messages. GitHub Copilot's June pricing pivot is the same story from the vendor side.

Apr 27, 20266 min read

Operating AI

Sierra Hires Differently. Meta Cuts Differently. 95% of AI Pilots Still Fail.

Hiring redesign, workforce cuts, and ROI failure data are three sides of the same restructure. Pick one without the others, and you stay on the failure curve.

Apr 27, 20266 min read

Operating AI

When Claude Is the Price Floor: A CFO's Renewal Math

A CFO replaced a vendor's AI feature with Claude integrations: 95% of the work, 15% of the cost, 45% spending cut. Vendors now sell what Claude can't replicate.

Apr 27, 20265 min read

Operating AI

Marketing Just Got Its First Engineering Outage. Nobody Wrote the Postmortem.

Growth Engineers run agent fleets that scrape competitors, generate 50 ad variations, and pause campaigns daily. None of the SRE practices exist yet.

Apr 27, 20266 min read

The AI Control Problem

Postmortem Culture Just Reached AI

Anthropic's April 23 postmortem, Sysdig's 12h31m exploit timeline, and a layered defense for AI-generated UI all point one direction: AI is becoming SRE.

Apr 27, 20267 min read

Operating AI

When You Certify an Agent, You Change What It Optimizes For

A peer-reviewed eBay study (JMR 2025) shows sellers cluster above certification cutoffs. The same mechanism is now built into every AI eval gate.

Apr 27, 20266 min read

Governed Implementation

Three Hyperscalers Shipped the Same Idea in One Week. Workspace Is the Control Plane.

OpenAI, Google, and Adobe each shipped a workspace-as-agent-environment play in seven days. The platform layer for agent governance is being defined right now.

Apr 27, 20266 min read

Operating AI

Cloudflare Just Published Its Own Case Study

Cloudflare dogfooded the governance stack it sells. 93% R&D adoption, MR velocity 5,600 to 8,700/week. Read it for the pattern, not the numbers.

Apr 23, 20266 min read

Operating AI

AI Industrializes Coordination. Not All of It.

A pseudonymous CTO replaced half a dozen management seats with a LangGraph mesh. The interesting part is what stayed human, not what left.

Apr 23, 20267 min read

Operating AI

HR Has an AI Governance Surface Now. McKinsey Just Gave It Numbers.

McKinsey quantified the people-side of AI adoption. That turns HR, manager behavior, and encouragement signals into an auditable governance surface.

Apr 23, 20268 min read

Operating AI

Marketing's Governance Stack Is Real. It's Also Asymmetric.

Three marketing governance surfaces showed up this month. Only one is actually a governance layer. Naming the asymmetry tells you where to invest.

Apr 23, 20267 min read

Operating AI

Four Reports, One Week, One Measurement Gap

Four independent reports read together expose a structural pattern. Organizations claim AI usage but cannot count it on a shared scoreboard.

Apr 23, 20267 min read

The AI Control Problem

Governance Goes on Stage: What UKG's SOC Math Reveals About AI Operations in 2026

A Fortune-scale customer just published its AI-agent SOC unit economics from a Google Cloud Next stage. What that signals, and what it doesn't.

Apr 23, 20268 min read

Governed Implementation

The Ungoverned 60%: Why Context Engineering Is Becoming a Cross-Functional Discipline

Capability is no longer the bottleneck. The ungoverned 60% of product knowledge outside the codebase is. Three writers converge on the same fix.

Apr 23, 20268 min read

Operating AI

Your AI Agent Reads Your Code. It Doesn't Know Your Product.

Code is syntax. Product is judgment. The context your agent needs lives outside the repo, and most teams have never written it down.

Apr 21, 20266 min read

Operating AI

AI Economics Just Fractured on Three Axes in One Week

Cost per task rising. Usage growing faster than efficiency. Revenue inflated by accounting. The cheaper-and-better consensus broke this week.

Apr 21, 20268 min read

The AI Control Problem

The App Store Became the First Governance Chokepoint for AI-Generated Software

App Store releases jumped 60% YoY in Q1 2026 and 104% in April. Apple's review process is now the first large-scale governance chokepoint for AI software.

Apr 21, 20266 min read

Engineering Notes

A Top Agent's Core Is a While-Loop

Claude Code's architecture is a loop. The interesting work, and the governance surface, lives at the perimeter: tools, prompts, checkpoints.

Apr 21, 20265 min read

Operating AI

Claude Design Shipped. Linear's CEO Fired Back. Both Are Right.

Anthropic launched Claude Design on Thursday. Linear's Karri Saarinen published the rebuttal on Friday. Design just had its governance moment.

Apr 21, 20267 min read

Engineering Notes

What Changed in the Opus 4.7 System Prompt Is a Governance Reference

Simon Willison's 4.6-to-4.7 diff turns a vendor system prompt into an auditable reference for teams building their own AI guardrails.

Apr 21, 20265 min read

Operating AI

Cloudflare Just Shipped the Operating System for the Agent Web

Cloudflare shipped Readiness Score, Artifacts, and Flagship on the same day. The pattern is bigger than the products: governance is now product.

Apr 21, 20267 min read

Governed Implementation

The Bounding Box Comes From Infrastructure, Not From a Prompt

Docker's microVM architecture makes the argument concrete. An LLM deciding its own security boundaries is not a security model. Infrastructure is.

Apr 21, 20266 min read

Operating AI

Machine-Readability Is the CMO's New KPI

Three signals in one week converged on a single truth. Marketing now owns a governance surface it has never had to think about: machine-readability.

Apr 21, 20267 min read

Engineering Notes

The Three-Layer Connectivity Stack Just Became Official

Anthropic's MCP co-creator just made the 2026 agent stack official. Here is what procurement should demand before Q2 contracts close.

Apr 21, 20268 min read

Operating AI

What Netflix's Live-Ops Playbook Teaches About Operating Agent Fleets

Netflix went from 1 live event a month to 400+. The lesson for AI isn't about automation. It's about the human operations layer that scales alongside it.

Apr 21, 20266 min read

The AI Control Problem

Shadow AI Is the New Supply Chain. Vercel Just Proved It.

Vercel's April breach is the first public production incident where the vector was ungoverned AI tool adoption. The fix is a company brain.

Apr 21, 20267 min read

Governed Implementation

Your Agent Permission Model Works 40% of the Time

ManyIH-Bench shows frontier models resolve instruction hierarchy at ~40% when sources exceed four tiers. Your permission model runs on a 40% base.

Apr 20, 20268 min read

The AI Control Problem

The Year Offensive AI Started Rewriting Open Source

Cal.com went closed source after five years. Codex rooted a Samsung TV. A one-line config blocks half of historical npm attacks. The pattern matters.

Apr 20, 20269 min read

Operating AI

Credit Governance Is the Template for AI Agent Spend

Four credit models from OpenAI, Cursor, Clay, and Vercel show how enterprises should govern AI agent spend before shadow credit burns the budget.

Apr 20, 20269 min read

Operating AI

Enterprise MCP Just Crossed the Engineering Line

Cloudflare documented MCP rollout across product, sales, marketing, finance with named governance controls. Here is what that means for buyers.

Apr 20, 20269 min read

The AI Control Problem

When the Harness Changes and the Model Does Not

Anthropic did not nerf Claude. The harness changed. Here is the five-surface audit framework enterprise buyers should demand.

Apr 20, 20269 min read

Operating AI

Your Marketing Team Just Became a Governance Team

Six signals in one week, including +80.4 percentage points from a single product rename, say marketing is now a governance job without an owner.

Apr 20, 202610 min read

The AI Control Problem

The Week Probabilistic Engineering Became Observable

Tim Davis named probabilistic engineering on April 16. The same day, Meta and monday.com shipped the proof it scales. Three signals, one operating model.

Apr 20, 20269 min read

The AI Control Problem

The Fifth Theme the Thoughtworks Radar Missed

Radar v34 named four governance themes for engineering. The same patterns are already loose in marketing, legal, HR, sales, and finance.

Apr 20, 202610 min read

Operating AI

The Workflow Is the Governance Primitive

Jakob Nielsen (UX Tigers) says AI's design problem is organizational. Victorino extends it: the workflow is the governance primitive.

Apr 20, 20268 min read

Governed Implementation

OpenAI's Agents SDK Learned to Run. It Still Cannot Be Governed.

OpenAI's SDK update adds sandboxes and memory. It solves execution. It does not solve governance. The attack surface grew; the controls did not.

Apr 16, 20268 min read

Engineering Notes

Screenshots Are Lazy Context Engineering

Callstack cut agent token usage by 50% without changing models. They just stopped sending lazy context. Cost governance hiding as an engineering note.

Apr 15, 20265 min read

Operating AI

Four Launches, One Week, One Control Taxonomy

Cloudflare shipped identity, network, cost, and coordination controls for AI agents in a single week. The governance baseline moved.

Apr 15, 20266 min read

Operating AI

Cost Governance in the Flat-Fee Era: The 3-Layer Framework You Should Have Shipped Yesterday

The flat-fee era of enterprise AI ended this quarter. Cost governance is now a three-layer problem buyers must solve before the board asks why.

Apr 15, 20267 min read

Operating AI

Engineering Has Cloudflare. Marketing Has Nothing.

Engineering built a real AI governance stack in twelve months. Marketing, design, content, and sales are still operating agents with no control surface.

Apr 15, 20266 min read

Engineering Notes

Cursor's Multi-Agent Kernels: What Measured Agent Teams Look Like

Cursor shipped 235 optimized CUDA kernels with a 38% average speedup. The real story isn't the number. It's the measurement discipline behind it.

Apr 15, 20265 min read

Operating AI

The Toggle Is the Point: Google Just Made Human Review a UI Element

Google shipped a 'Require human review' toggle inside its desktop Agent. It's a small UI element. It's also a governance precedent.

Apr 15, 20264 min read

The AI Control Problem

OpenAI Just Became a Compliance Actor

GPT-5.4-Cyber is a governance precedent dressed as a cybersecurity launch. Your vendor is now segmenting your team. Read it twice.

Apr 15, 20264 min read

Operating AI

Team Alignment Is the New Bottleneck

One developer, two dozen agents, zero alignment. Individual velocity is not the moat anymore. Team coherence at agent-scale is.

Apr 15, 20268 min read

Operating AI

Trust Is the UX

AI polished everything. So polish stopped being a trust signal. Now imperfection, strategic friction, and verifiable metrics do the work instead.

Apr 15, 20266 min read

The AI Control Problem

The Week Verification Became Doctrine

Verification got a manifesto. Reproducibility got a paper. Cross-model review shipped as a CLI feature. The argument is over. The doctrine is here.

Apr 15, 20266 min read

Operating AI

Your Docs Have Two Audiences Now. One of Them Counts Tokens.

Addy Osmani named it Agentic Engine Optimization. Your information architecture is now a governance surface for AI agents, not just humans.

Apr 13, 20267 min read

The AI Control Problem

Capability Is a Commodity. Scaffolding Is the Moat.

Three April 2026 papers dismantle the 'which model should we buy' question. The real moat is the harness, not the model.

Apr 13, 20269 min read

Operating AI

Datadog Just Turned Governance Into a Product Roadmap

Two Datadog AI-governance releases in one week — a Code Security MCP and an AI-native SAST. That's not compliance. That's a product category.

Apr 13, 20266 min read

Governed Implementation

When Designers Code Again, the Org Chart Moves

LukeW says designers should code again. The real story is governance - when AI collapses the handoff wall, the boundary moves and ownership shifts.

Apr 13, 20266 min read

The AI Control Problem

The $7 Doritos Moment: Enterprises Are Repricing AI as Discretionary

48% of executives now call AI a massive disappointment. The industry is being repriced as discretionary. Measurement is the moat.

Apr 13, 20267 min read

Operating AI

The Accountability Asymmetry: Why Agents Buy Small and Humans Still Buy Big

Agents can pick software. They can't face your board. That single asymmetry is redrawing the governance boundary of every GTM motion.

Apr 13, 20266 min read

The AI Control Problem

Laziness Is a Virtue. Your LLM Doesn't Have It.

Good engineers are lazy. LLMs structurally can't be. If you measure AI by lines produced, you are rewarding the opposite of good engineering.

Apr 13, 20266 min read

Operating AI

Marketing Just Had Engineering's 2024 Moment

Four April 2026 signals put marketing ops where engineering was 18 months ago: budget, compounding, authority, shadow AI. The fabric transfers.

Apr 13, 20268 min read

Governed Implementation

Orchestration Is the New Governance Layer

Four April 2026 signals converge: role separation — planner, worker, validator — is replacing prompt engineering as the unit of AI governance.

Apr 13, 20269 min read

Operating AI

Execution Speed Was the Moat. Learning Speed Is.

AI collapses translation costs between departments. The hierarchy those costs justified is the next thing to collapse.

Apr 13, 20267 min read

Governed Implementation

Your Harness, Your Memory — And the Moat You Can't Take With You

If memory is the moat and memory lives in the harness, a vendor-owned harness is a vendor-owned moat. The exit question most teams can't answer.

Apr 11, 20267 min read

Operating AI

Position One Is the New Page One: Inside Google AI Mode Shopping

93% of shoppers say they double-check AI. Only 5% actually do. The stated-vs-revealed gap is a marketing governance problem.

Apr 10, 20266 min read

Governed Implementation

Claude Managed Agents: When the Harness Becomes a Vendor Product

Anthropic turned the agent harness into a SKU. Build-vs-rent is now a governance decision, not a technical one. Here is the split.

Apr 10, 20267 min read

Operating AI

Designers as Conductors: AI Is Rewriting Design's Operating Model

AI didn't invent the senior-IC orchestration path. It re-timed it, pulling staff-level skills down into the middle of every design team.

Apr 10, 20266 min read

Governed Implementation

The Manifesto Arrives: Auditing McKinsey's 12 Themes

McKinsey names trust and guardrails for the first time. The sixth chapter in our tracking series. What the manifesto still leaves out.

Apr 10, 20268 min read

The AI Control Problem

The DHH Setup Is Not a Tool Stack. It Is an Org Chart.

Productive AI use is an organizational shape, not a tool choice. Before you copy DHH's workflow, copy the preconditions that make it work.

Apr 10, 20267 min read

The AI Control Problem

Post-SaaS Economics: The Premium Is Gone. Discipline Is the New Multiple.

Software now trades at a discount to the S&P 500 for the first time in the SaaS era. What CFOs and boards should do about it.

Apr 10, 20267 min read

The AI Control Problem

This Is Not the Dot-Com Bubble. It Is Also Not a Free Lunch.

What Fowler and Beck teach about AI adoption. Two truths must coexist: the technology is real, and the industry around it is replaying old mistakes.

Apr 10, 20269 min read

The AI Control Problem

The Plugin That Wrote Its Own Consent Dialog

Vercel's Claude Code plugin asks permission through an injected system prompt. You cannot tell who wrote the question. That is the real problem.

Apr 10, 20267 min read

Governed Implementation

The Agent Production Stack Now Has Governance Built In. But Whose Governance?

Anthropic decouples brain/hands/session. Claw-Eval grades safety alongside completion. Infrastructure catches up to capability.

Apr 9, 20269 min read

The AI Control Problem

29% of Fortune 500 Pay for AI. Governance Isn't Blocking Adoption — It's Shaping It.

a16z data shows coding dominates AI adoption by an order of magnitude. Regulated industries lead. No single vendor wins.

Apr 9, 20268 min read

Engineering Notes

AI Code Review Is Setting Its Own Standards. Who Reviews the Reviewer?

BugBot hits 78% resolution with 44K autonomous rules. Claude PR review needs human-written config. Same question, two answers.

Apr 9, 20269 min read

Operating AI

164 Million Purchases Exposed AI Traffic's Conversion Problem

AI referrals convert 11.5% worse than organic search across 973 sites. Marketing governance demands evidence over narrative.

Apr 9, 20266 min read

The AI Control Problem

Contained Financial Harm vs. Active Military Conflict: The Appeals Court Frames AI Governance

A D.C. court denied Anthropic's stay. The framing — company harm vs. national security — crystallizes AI's governance tension.

Apr 9, 20268 min read

Operating AI

Your Content Moderator Explains Itself Fluently. Its Explanations Are Wrong.

Google's LLM moderation survey reveals a governance crisis: models generate plausible rationales disconnected from their actual decisions.

Apr 9, 20269 min read

Operating AI

99.5% AI Adoption at a $32B Company. The Secret Wasn't the Technology.

Ramp's CPO reveals an L0-L3 proficiency ladder and 6,300% usage growth. AI adoption is an org design problem.

Apr 9, 20268 min read

Operating AI

Your Agent Forgets Everything. Here Are Three Ways to Fix It — and One Question Nobody Is Answering

IBM's ALTK-Evolve improves hard tasks by 74%. Fowler's Flywheel compounds team AI practice. Neither solves who validates what agents learn.

Apr 9, 202610 min read

The AI Control Problem

AI Reads Text. It Guesses Charts.

Frontier models lose 16-20 points of accuracy reading charts vs text. GitHub lost a nine to agent traffic. The competence wall is not theoretical.

Apr 8, 202610 min read

The AI Control Problem

Project Glasswing: When the Lab Won't Release Its Own Model

Anthropic built a model too capable for general release. That decision tells you more about AI governance than any framework document.

Apr 8, 20269 min read

Operating AI

Your AI Fabricated 30 Prospects. Marketing Has a Governance Problem.

A veteran marketer documents LLMs fabricating data, inventing metrics, and lying about verification. Meanwhile, Meta wants full ad automation by year-end.

Apr 8, 20269 min read

Operating AI

671 PRs, Zero Reverts: The Verification Revolution Has Data Now

Vercel auto-approved 671 low-risk PRs with zero reverts. GitHub's cross-model review closes 74.7% of performance gaps. The theory phase is over.

Apr 8, 20269 min read

Operating AI

The Agent Operations Stack Is Shipping

A 28-test reliability framework, centralized guardrails from AWS, and a 3-human company running 20 agents. The infrastructure for agent operations just arrived.

Apr 6, 20269 min read

Operating AI

AEO Is Already Commoditized. The Durable Play Is Governing What AI Trusts.

First-wave AI search optimization is a Red Queen race. The organizations that win will govern hard-to-fake trust signals, not chase formatting tactics.

Apr 6, 20268 min read

Engineering Notes

An AI Found Five Linux Kernel Bugs. Now What?

A researcher used Claude to find a 23-year-old Linux vulnerability. The governance question is not whether AI can find bugs. It is who else is already looking.

Apr 6, 20268 min read

Operating AI

Make Kits Shipped. Sora Collapsed. The Lesson Is the Same.

Figma's Make Kits constrain AI to production components. Sora died without that constraint. Design systems are governance infrastructure.

Apr 6, 20268 min read

The AI Control Problem

Your AI Decides Before It Thinks

New research shows LLMs encode action decisions before generating reasoning tokens. Chain-of-thought may be rationalization, not deliberation.

Apr 6, 20267 min read

The AI Control Problem

MEDVi: $400M in Revenue, Two Employees, and the Healthcare Governance Vacuum

MEDVi built $400M in revenue on rented clinical infrastructure and AI wrapping paper. The real story is what healthcare governance failed to prevent.

Apr 6, 202610 min read

The AI Control Problem

The Mercor Breach Exposed AI's Most Guarded Secret: How Models Get Trained

A supply chain attack on Mercor exposed proprietary training data for Meta, OpenAI, and Anthropic. Secrecy is not security.

Apr 6, 20268 min read

Engineering Notes

When the Harness Engineers Itself

Stanford's Meta-Harness automates harness optimization with an agentic search loop. If harnesses are governance, governance is now optimizable by machines.

Apr 6, 20267 min read

The AI Control Problem

Software Slop Is an Attention Problem

Slop is not bad code. It is code nobody looked at carefully. A new metric quantifies the missing governance signal: attention deficit.

Apr 6, 20267 min read

Operating AI

Your Agent Skips Its Own Memory. That Is a Governance Failure.

Agents default to convenience over governed retrieval — silently ignoring tool-based memory systems in favor of flat files always in context.

Apr 4, 20268 min read

Governed Implementation

SDD at Enterprise Scale: A Governance Problem in Tooling Clothes

Spec-driven development's 7 enterprise adoption barriers are organizational, not technical. Here is what the tooling vendors won't tell you.

Apr 4, 202611 min read

Operating AI

Governed Agent Infrastructure at Scale: What KernelEvolve and AWS DevOps Agent Actually Prove

Agent infrastructure is shipping with governance built in. Meta and AWS show what evaluation-driven autonomy looks like at production scale.

Apr 4, 20268 min read

The AI Control Problem

The Noise Is the Signal: METR Data Shows AI Measurement Getting Harder, Not Easier

New METR data: benchmarks saturating, task noise spanning 4x ranges, human baselines costing $8K+. Measurement itself is now the bottleneck.

Apr 4, 20269 min read

The AI Control Problem

The Week AI Monitoring Failed at Every Layer

Three independent findings converge: AI models protect peers, RL training hides reasoning, and commercial decisions erase transparency.

Apr 4, 202610 min read

Operating AI

Open Models Crossed the Agent Threshold. Now What?

Open-weight models now match frontier systems on core agent tasks. Choosing the expensive option is a governance decision, not a capability requirement.

Apr 4, 20268 min read

Operating AI

Platform Coupling: Why AI Cites What It Owns, Not What You Publish

45.2 million citations reveal AI search is shaped by licensing deals, not content quality. Grok cites X 99.7%. ChatGPT favors Reddit. Governance implications.

Apr 4, 20268 min read

Operating AI

Silent Drift: The Agent Failure Mode Nobody Named

AI code that compiles, passes tests, and quietly violates architectural assumptions. Silent drift is the agent operations risk nobody is measuring.

Apr 4, 202610 min read

The AI Control Problem

Your AI Was Designed to Agree With You

Leaked Gemini directives reveal AI systems hardcoded to validate emotions over accuracy. Trust engineering demands governance, not agreement.

Apr 4, 20268 min read

The AI Control Problem

Your LLM Benchmark Score Is a Scaffold Artifact. Here Is What Actually Matters.

Switching evaluation scaffolds causes 15% performance swings. IRT cuts eval costs 140x. The benchmark is not broken — your measurement infrastructure is.

Apr 2, 20269 min read

Governed Implementation

Structured Prompting Is a Governance Mechanism for AI Code Reasoning

Semi-formal reasoning forces AI agents to show verifiable evidence before making code judgments. The real value is auditability, not accuracy.

Apr 2, 20269 min read

The AI Control Problem

There Is No Universal AI Citation Formula

98K citation rows across 7 verticals reveal that signals lifting one industry's AI visibility actively hurt another.

Apr 1, 20269 min read

The AI Control Problem

NVIDIA Takes 79% of AI Profit. Applications Get 7%. The Value Chain Won't Self-Correct.

Two years of data confirm AI's inverted value chain. Semiconductors capture 79% of profit. The concentration risk is a governance problem.

Apr 1, 202610 min read

Operating AI

The Infrastructure Between Agent Demos and Agent Operations

70% of Devin sessions are human-triggered. Bessemer maps five infrastructure frontiers where governance primitives don't yet exist.

Apr 1, 20269 min read

The AI Control Problem

The App Store Just Became AI's First Distribution Checkpoint

Apple is pulling vibe-coded apps that generate and run code outside review. The governance framework was already there. It just needed enforcement.

Apr 1, 20266 min read

The AI Control Problem

The Autopilot Reckoning: When AI Sells the Work, Who Governs the Output?

Sequoia says the next $1T company sells AI work, not tools. JustPaid runs 7 agents for $15K/month. The governance vacuum is widening.

Apr 1, 202612 min read

Engineering Notes

Only 27% of ChatGPT Citations Rank on Google. That Is a Governance Problem.

New data: 60% of ChatGPT citations don't appear in search results at all. Organizations can't govern AI visibility like SEO.

Apr 1, 20269 min read

The AI Control Problem

512,000 Lines of 'Safety-First': What Claude Code's Source Leak Reveals About AI Governance Theater

Anthropic leaked Claude Code's full source to npm. The code shows anti-distillation tricks, stealth modes, and bypassed permissions. Two leaks in five days.

Apr 1, 202612 min read

Operating AI

From Tribal Knowledge to Governed Intelligence: Meta Runs 50,000 Daily Analyses

Meta's DrP platform codifies debugging expertise into testable analyzers. 300 teams, 20-80% MTTR reduction. The knowledge capture pattern is governance.

Apr 1, 202610 min read

The AI Control Problem

Every Structural Governance Attempt for AI Labs Has Failed. The DeepMind Files Explain Why.

DeepMind's 3-3-3 board, $1B walk-away plan, nonprofit shell — all blocked. Mallaby's account reveals why AI governance keeps losing to power.

Apr 1, 202614 min read

Operating AI

The Design Twin Problem: Who Governs What AI Builds?

AI compresses design's production layer, exposing a discipline narrowed to UI. The Design Twin concept demands governance nobody has built yet.

Apr 1, 20268 min read

Governed Implementation

Thoughtworks Discovered Governance-as-Code. We Have Been Building It.

Fowler's team frames AI instruction files as developer tooling. The real move is organizational governance. Here is where their pattern stops and ours begins.

Apr 1, 20268 min read

Operating AI

The Week Governance Became a Product Feature

Anthropic ships audit logs. Microsoft ships dual-model critique. Both are governance primitives embedded in product, not bolted on.

Apr 1, 20265 min read

Operating AI

AI Governance Is Leaving the Engineering Silo

Adobe ships brand governance as product. Dorsey replaces management with AI. A marketing newsletter uses CI/CD vocabulary. The pattern is clear.

Apr 1, 202611 min read

Governed Implementation

The Intent Layer: Why the Best AI Governance Is the Kind Nobody Notices

Spacelift Intent, Google Skills, Claude Code harnesses — the pattern is clear. Effective governance is invisible, embedded in tooling, not bolted on.

Apr 1, 20268 min read

The AI Control Problem

The First AI Gateway Supply Chain Attack: LiteLLM Breach Hits a $10B Startup and Thousands More

LiteLLM compromise intercepted prompts, responses, and API keys at scale. Mercor ($10B) was one of thousands affected. AI supply chain risk is real.

Apr 1, 20269 min read

Operating AI

Mandatory Agent Review: What VS Code's 2.2x Commit Growth Actually Required

VS Code doubled commits by making Copilot review mandatory. Nango built 200+ integrations for under $20. Both required governance first.

Apr 1, 20269 min read

Governed Implementation

Making Your Product Agent-Ready Is a Governance Decision

AI agents are selecting vendors autonomously. Optimizing for agent discovery without governing what you expose is a liability.

Apr 1, 20267 min read

Governed Implementation

The Spec Layer: Five Independent Teams Discovered the Same Agent Governance Architecture in One Week

Matt Rickard, Google, GitHub, Anthropic, and Kent Beck converge on specs as the primary governance mechanism for AI agents.

Apr 1, 202614 min read

The AI Control Problem

The Two-Front Supply Chain Crisis: AI Is Breaking Open Source From Both Sides

AI agents find vulnerabilities faster than defenders can patch. AI-generated contributions overwhelm maintainers. Both hit the same 96% dependency surface.

Apr 1, 20269 min read

Engineering Notes

Workload-Harness Fit: The Governance Framework Hiding in Agent Lab Economics

Agent labs are training their own models. Whether to train vs. harness depends on four dimensions that double as a governance framework.

Apr 1, 20269 min read

Governed Implementation

Your AI Will Hallucinate. Build the System That Catches It.

Hallucinations are intrinsic to probabilistic AI. The discipline that matters is not elimination but governance: layered detection, validation, and control.

Mar 31, 20269 min read

Governed Implementation

OpenAI Built Governance Into the Platform. That Is Not Enough.

OpenAI built governance into their agent platform. That solves the containment problem. It does not solve the governance problem.

Mar 31, 202610 min read

Operating AI

The Two Governance Problems in Agentic Commerce

AI agents now choose software and spend tokens autonomously. One problem is who governs what they buy. The other is who governs what they cost.

Mar 30, 20268 min read

The AI Control Problem

Your AI-Generated Code May Already Be Public Domain

The U.S. Supreme Court declined to hear the AI copyright appeal. Code written by agents cannot be copyrighted. Most teams have no plan for this.

Mar 30, 20268 min read

The AI Control Problem

Your AI Says Yes Too Often. Science Proves It Changes Who You Are.

Stanford study in Science: LLMs affirm users 49% more than humans and endorse harmful behavior 47% of the time. Sycophancy is now a governance problem.

Mar 30, 20269 min read

The AI Control Problem

The Cognitive Dark Forest and the Case for Knowledge Governance

When every public idea becomes training data, sharing innovation is a strategic exposure. The answer is not secrecy. It is deliberate knowledge governance.

Mar 30, 20266 min read

Operating AI

Design Debt in AI Products: When Interface Decisions Shape What Users Believe

Tech debt is tracked and budgeted. Design debt is not. In AI products, that silence is dangerous because design shapes belief, not just behavior.

Mar 30, 20267 min read

The AI Control Problem

The Determinism Divide: Why Executives and ICs Need Different AI Governance

Executives manage chaos by default. ICs are measured on precision. AI governance that ignores this split will fail both groups.

Mar 30, 20267 min read

Governed Implementation

Free Software in the Agent Era: A Governance Reckoning

AI agents let non-technical users modify and redistribute software. Open-source is now a governance decision, not just a technical one.

Mar 30, 20269 min read

Engineering Notes

From 6.75% to 99.8%: What Type-Constrained Verification Delivers

A type-constrained harness turned Qwen's 6.75% function calling success into 99.8%. The harness tax is negative. It pays for itself on the first call.

Mar 30, 20268 min read

Operating AI

ISO 42001: When AI Governance Becomes a Product Feature

Datadog's ISO 42001 certification signals a shift. AI governance is no longer internal discipline. It is becoming a vendor selection criterion.

Mar 30, 20267 min read

Operating AI

Platform Governance Is Splitting: Reddit Labels Bots While ChatGPT Sells Them Ad Space

Reddit and ChatGPT chose opposite governance models for AI presence. Both models carry brand risk. The question is which risk you are managing.

Mar 30, 20267 min read

Operating AI

Just One More Prompt: The Dopamine Trap of AI-Assisted Work

BCG found AI brain fry affects 14% of workers. 39% more major errors. The mechanism is the same one that makes slot machines work.

Mar 28, 20269 min read

Governed Implementation

McKinsey's Two Paths Are One Question They Won't Ask

McKinsey frames enterprise architecture as incremental vs comprehensive. The real choice is governance maturity. Fifth in our tracking series.

Mar 28, 20269 min read

Engineering Notes

What Is an Agentic Mesh? The Infrastructure Layer Your Multi-Agent System Is Missing

An agentic mesh connects isolated AI agents into a governed network. Most organizations need one. Few know it exists.

Mar 28, 202610 min read

Operating AI

Advertising Discovers Governance. Two Years Late.

AI ad placement now faces the same governance deficit engineering solved years ago. But advertising has less infrastructure and higher stakes.

Mar 27, 20268 min read

The AI Control Problem

The First Amendment as Governance Shield: What the Anthropic Ruling Changes

A federal judge ruled the Pentagon's Anthropic ban is illegal First Amendment retaliation. AI governance red lines now have judicial backing.

Mar 27, 20268 min read

The AI Control Problem

LiteLLM: When the AI Gateway Becomes the Attack Vector

A malicious PyPI update to LiteLLM exfiltrated SSH keys, cloud credentials, and K8s tokens. Your AI middleware is your attack surface.

Mar 27, 20269 min read

Operating AI

When AI Mandates Become Org Charts: Meta's Structural Bet

Meta's leaked AI targets are less interesting than its org restructuring. Renaming employees 'AI Builders' encodes identity, not just policy.

Mar 27, 20269 min read

Governed Implementation

Your Governance Reviews Quarterly. Your Models Update Every 5 Hours.

Cursor ships model checkpoints every 5 hours. OpenAI's own training lags its spec. Governance built for quarterly review cannot survive continuous deployment.

Mar 27, 20269 min read

Engineering Notes

Seven Hours to Generate, One Week to Trust

Reco.ai spent $400 and 7 hours on AI code generation. Verification took a full week. Three sources confirm: judgment is the last constraint.

Mar 27, 20269 min read

Operating AI

The 400-Dollar AI Team That Nobody Governs

A developer built a 24/7 AI agent team for 400 dollars a month. The patterns are real. The missing governance makes it a liability.

Mar 26, 20269 min read

The AI Control Problem

When the Security Gate Becomes the Vulnerability: Claude Code's Auto Mode and the Classifier Gambit

Anthropic's data shows 93% of permission prompts get rubber-stamped. Auto mode replaces human approval with a model classifier. The tradeoffs are real.

Mar 26, 20269 min read

Governed Implementation

Design-First Is the Cheapest AI Governance You Can Buy

AI assistants collapse design into code. A five-level framework forces the decisions back into the open. But agentic AI needs more.

Mar 26, 20268 min read

Governed Implementation

Design Systems Just Became AI Governance Infrastructure

Figma's MCP beta lets agents write to the canvas. Design systems are no longer style guides. They are constraint layers for autonomous software.

Mar 26, 20266 min read

The AI Control Problem

25,000 Agents, 100,000 Lawyers, Zero Governance Standards

Harvey's $11B valuation proves vertical AI agents are real. The absence of governance standards proves nobody is ready for what comes next.

Mar 26, 20267 min read

Engineering Notes

When Your AI Explains Its Reasoning, It's Making It Up

Anthropic's interpretability research shows Claude uses different computation than it describes. Chain-of-thought is post-hoc rationalization.

Mar 26, 20267 min read

Operating AI

Your AI Agent's Judgment Is Unmeasured. A PM Rubric Shows What to Fix.

Jeff Gothelf's 4-dimension rubric for product judgment applies directly to AI agents. Nobody is scoring agent decisions. Here is how to start.

Mar 26, 20268 min read

Operating AI

Marketing Agent Governance: What Klaviyo's Composer Reveals

Klaviyo shipped autonomous marketing agents with explicit governance controls. The governance question is no longer confined to engineering.

Mar 26, 20266 min read

Governed Implementation

McKinsey's Agent Factory Prescription Is Missing Its Own Diagnosis

McKinsey prescribes agent factories to fix 80% AI failure. They left out governance. Again. The fourth chapter in a pattern we have been tracking.

Mar 26, 20268 min read

The AI Control Problem

The Monetizable Spread: Why Open-Source Compression Makes Governance the Last Moat

Open-source AI trails frontier models by 3 months, not 12. The revenue zone is compressing. Governance is what remains.

Mar 26, 20267 min read

Governed Implementation

OpenAI's Governance Infrastructure Goes Public

OpenAI published its Model Spec and Safety Bug Bounty. What these artifacts reveal about treating governance as product infrastructure.

Mar 26, 20269 min read

Operating AI

The Simulation Toolchain Arrives. The Validation Infrastructure Does Not.

Open-source tools now let organizations simulate futures with thousands of AI agents. The danger is not inaccuracy. It is how convincing wrong answers become.

Mar 26, 202610 min read

The AI Control Problem

The Four Failure Modes of Ungoverned AI Coding

A practitioner's taxonomy of how AI agents decay codebases. Each failure mode maps to a governance control you can implement today.

Mar 26, 20268 min read

Engineering Notes

What Is an Agent Harness? The Concept Your AI Strategy Is Missing

An agent harness is the infrastructure wrapping your AI model. Same model, different harness, 85% better results. Here is what that means.

Mar 26, 20269 min read

Operating AI

Your Security Scanner Was the Attack: What Syscall-Level Monitoring Reveals About Agent Runtime

Sysdig captured 64 process executions in one AI session. Trivy was backdoored for 4 days. Application-level security cannot see either.

Mar 25, 20269 min read

The AI Control Problem

When Governance Has a Price Tag: What the Anthropic Court Challenge Reveals

A federal judge says the US government appears to be punishing Anthropic. The supply-chain designation was reserved for Chinese entities.

Mar 25, 20268 min read

Operating AI

Evaluation-Driven Development: The Missing Operations Layer for AI Agents

monday.com found 40% of agent failures were tool parameter errors. ServiceNow proved agents trade accuracy for experience. Evaluation is governance.

Mar 25, 20269 min read

Governed Implementation

Three Roads to Governed Autonomy: What PostHog, Ossature, and Anthropic Converge On

PostHog built 44 tools before choosing MCP. Ossature added verification loops. Anthropic shipped a safety classifier. All constrain the environment.

Mar 25, 20269 min read

The AI Control Problem

The Learning Curve Tax: What Anthropic's Own Data Reveals About AI Productivity

High-tenure Claude users get 10% better results. Average task value is declining. PyPI shows zero production surge. The productivity gap is governance.

Mar 25, 202610 min read

The AI Control Problem

The Collina Paradox: When the AI Evangelist Triggers the Backlash

A Node.js core contributor's AI-assisted PR triggered a 90+ signatory petition to ban AI code. The paradox reveals a governance vacuum, not hypocrisy.

Mar 24, 202610 min read

Engineering Notes

Generator-Evaluator Loops: What Anthropic's Harness Design Actually Teaches

Anthropic's long-running app harness reveals generator-evaluator loops, sprint contracts, and context decay. The data is thin. The patterns are real.

Mar 24, 202610 min read

The AI Control Problem

The Diagnosis Is Right. The Cure Doesn't Exist Yet.

Tsinghua researchers mapped every way an AI agent can be attacked. Their proposed defenses are theoretical. That honesty is the most useful part.

Mar 24, 202610 min read

Engineering Notes

Your 'Expert AI' Prompt Is Making Your AI Dumber

USC research: expert persona prompts drop accuracy 3.6pp while boosting safety 17.7pp. The fix is not removing personas. It is routing them.

Mar 24, 20269 min read

The AI Control Problem

The 2% Problem: What Federal Reserve Data Reveals About AI Productivity Theater

40% of workers use AI. 2% of hours are saved. The gap between adoption and impact is not a technology problem. It is a measurement problem.

Mar 24, 202611 min read

Operating AI

The Week Agent Infrastructure Went Mainstream

Four announcements in one week reveal the production agent stack: identity, orchestration, observability, and governance as first-class infrastructure.

Mar 23, 202610 min read

Operating AI

The Most Valuable Hire You're Not Making

A $6B company's AI agent quoted wrong pricing for a year. The missing role isn't technical. It's governance.

Mar 23, 20266 min read

The AI Control Problem

From 0% to 91%: Why Agent Safety Is Configuration-Dependent

An AI model's scheming rate jumps from near-zero to 91% with a prompt change. This has consequences for every governance framework built on static testing.

Mar 23, 202610 min read

Operating AI

You Cannot Govern What You Cannot Simulate

Enterprise world models are the missing governance layer. Without simulation, managing AI agents is guesswork at scale.

Mar 23, 20267 min read

The AI Control Problem

The AI Workforce Inflection: When Token Consumption Becomes a Performance Metric

Four signals from one week reveal AI crossing from tool to workforce. Governance frameworks built for the old model are already obsolete.

Mar 23, 202610 min read

The AI Control Problem

Walmart's 3x Conversion Gap: What Happens When AI Checkout Meets Real Customers

ChatGPT checkout converts 66% worse than Walmart.com. The world's largest retailer just proved where trust actually lives.

Mar 23, 20265 min read

The AI Control Problem

Waymo's 82% Safety Claim and the Conditions That Make It True

Waymo robotaxis crash more often than humans but cause fewer injuries. The conditions that explain this paradox matter more than the headline.

Mar 23, 20268 min read

The AI Control Problem

80,508 People Told Anthropic What They Want From AI. 14.7% Said Governance.

Anthropic's 81K-person study confirms what governance data already showed: benefits and harms coexist in the same people. That changes the policy math.

Mar 20, 202610 min read

Operating AI

What Tens of Millions of Agent Sessions Reveal About Misalignment

OpenAI monitored tens of millions of coding agent sessions. Less than 1% showed misalignment. The math still produces tens of thousands of incidents.

Mar 20, 202610 min read

The AI Control Problem

Clinejection: When an AI Agent Becomes the Attack Surface

A prompt injection in Cline's issue triage bot led to a supply chain compromise. Three composed weaknesses. One GitHub account required.

Mar 20, 202610 min read

Governed Implementation

Three Teams, One Pattern: What PM, Security, and Infrastructure All Built for Agent Governance

A PM platform, a security team, and an infra provider independently built governed AI agents. They converged on four identical patterns.

Mar 20, 20268 min read

Governed Implementation

When Visa Extends Your Agent Payment Protocol: What MPP Means for Governed Commerce

Stripe and Paradigm launched MPP with Visa, Mastercard, and both AI labs. The protocol is live. The governance is not.

Mar 20, 20268 min read

Engineering Notes

The Market Chose Governance: What Ramp's AI Spend Data Reveals About Brand Trust

Ramp's spend data shows Anthropic grew from 4% to 24.4% in one year, winning 70% of head-to-head matchups. Governance posture drives commercial outcomes.

Mar 20, 20267 min read

The AI Control Problem

The Speed Illusion Has Three Names Now

Three independent sources converge: AI speed without governance produces negative outcomes. The pattern echoes a 30-year electrification delay.

Mar 20, 20268 min read

Engineering Notes

The Context Crisis: Three Architecture Bets That Shrink the Agent's World to Make It Work

One team burned 72% of their context on tool definitions. Three independent solutions converge on the same principle: less is more.

Mar 19, 20269 min read

The AI Control Problem

Seven Requirements for Institutional AI: What Individual Productivity Cannot Buy

Hebbia's CEO names seven structural gaps between individual AI productivity and institutional value. The framework our thesis was missing.

Mar 19, 20269 min read

Operating AI

The Review Layer Paradox: When AI Governance Becomes the Bottleneck It Was Meant to Prevent

A 30-minute bug fix becomes a 12-week delivery with three review layers. Pennarun quantifies what most teams feel but cannot prove.

Mar 19, 20268 min read

The AI Control Problem

Shadow AI Is Not the Problem. Your Missing Governance Is.

78% of employees use unapproved AI at work. Blaming them is easier than admitting your organization never built the controls.

Mar 19, 202610 min read

The AI Control Problem

The Speed Trap: Why Faster AI Coding Makes Your Delivery Slower

Code writing is 20% of delivery. Optimizing it creates traffic jams, not productivity. Three sources converge on the same diagnosis.

Mar 19, 20269 min read

Operating AI

Stripe's Blueprints: How Deterministic Rails Make Agentic Code Safe at 1,300 PRs Per Week

Stripe reveals the architecture behind Minions: blueprints for hybrid workflows, Toolshed for 500 curated tools, and 10-second devboxes.

Mar 19, 20269 min read

Governed Implementation

Three Views on Governing AI Code: Types, Specs, and Verifiable Trust

Vint Cerf says trust is infrastructure. Lean 4 says types are proof. A Haskell expert says specs become code. All three are right.

Mar 19, 202610 min read

Engineering Notes

What Hundreds of Skills Taught Anthropic About Governing AI Agents

An Anthropic insider reveals the operational patterns from hundreds of production skills. The lessons challenge how most teams think about agent capabilities.

Mar 18, 202610 min read

Operating AI

Agent Memory Self-Design: Who Validates the Validator?

An agent redesigned its own memory and improved recall from 60% to 93% for $2. The breakthrough is real. The governance gap is bigger.

Mar 17, 20269 min read

Operating AI

Running AI Agents at Scale: What Stripe, Cloudflare, and Istio Teams Actually Do

Four operational primitives separate teams running agents in production from those still demoing. The data is in.

Mar 17, 202610 min read

The AI Control Problem

The Evidence Is In: AI Coding Agents Are Breaking Things

Amazon outages, Anthropic's own bugs, mandated adoption backlash. The evidence against ungoverned AI coding is no longer theoretical.

Mar 17, 20269 min read

The AI Control Problem

$200M Says Verification Is Infrastructure. The Hard Part Is What Comes Next.

Axiom raises $200M at $1.6B to prove AI code correct with Lean 4. The market validated our thesis. The specification problem remains unsolved.

Mar 17, 20269 min read

The AI Control Problem

When Infrastructure Ships Governance: What Cloudflare's Free AI Security Means

Cloudflare made AI endpoint discovery free for everyone. The signal: governance is no longer optional. It is becoming infrastructure.

Mar 17, 20268 min read

The AI Control Problem

When AI Builds and Breaks: The Widening Code Governance Gap

Amazon's Kiro caused a 13-hour AWS outage. SWE-bench shows 12+ months of stagnation. The gap between AI deployment velocity and verification is growing.

Mar 17, 202610 min read

Governed Implementation

The Containment Convergence: Why NVIDIA, Docker, and OpenAI All Shipped Agent Sandboxes This Week

Three companies shipped agent containment in one week. The pattern is identical: YAML policies, egress proxies, credential isolation.

Mar 17, 202612 min read

The AI Control Problem

Domain Expertise Still Wanted: The AI Trust Gap Is Not Closing

64% of devs use AI to learn. Only 1% trust it alone. Stack Overflow's 2026 data shows the verification tax is now a permanent operating cost.

Mar 17, 20269 min read

The AI Control Problem

Governance Gates Enterprise AI — Not Capability

Enterprise AI adoption is blocked by permissioning, sandboxing, and regulatory caution. Model capability is no longer the bottleneck.

Mar 17, 20268 min read

The AI Control Problem

The Institutional AI Gap: Why Your Team's AI Productivity Gains Aren't Showing Up in Revenue

Individual AI productivity is real. Institutional AI productivity is not. The 30-year electrification parallel explains why — and what to do about it.

Mar 17, 202610 min read

Governed Implementation

McKinsey's Three Diagnoses: How a Consulting Firm Slowly Discovers Governance

McKinsey went from measuring AI wrong to calling it a design problem to using the word governable. The pattern reveals more than any single article.

Mar 17, 20268 min read

Engineering Notes

MCP Is Dead; Long Live MCP — Why the CLI vs Protocol Debate Misses the Point

CLIs save tokens for individuals. MCP wins for enterprises. The real question is not which is better but who is asking.

Mar 17, 20269 min read

The AI Control Problem

The Verification Tax

Executives report saving 4.6 hours per week with AI. Workers spend 3.8 hours checking it. The net gain is 16 minutes. Someone is paying for the illusion.

Mar 16, 20268 min read

The AI Control Problem

46.5 Million Messages in 2 Hours: Why Agent Security Is an Architecture Problem

McKinsey's AI platform breached via basic SQL injection. OpenAI reframes defense as blast radius control. Security requires architecture, not prompting.

Mar 12, 202610 min read

The AI Control Problem

When Your Customer Is an Algorithm: The Governance Problem in Agentic Commerce

Three competing protocols. $385B at stake. Zero governance standards. The real moat in agentic commerce is not optimization.

Mar 12, 20269 min read

The AI Control Problem

Who Watches the 700 Experiments?

Karpathy's autoresearch runs hundreds of AI experiments overnight. The tool works. The governance does not exist.

Mar 12, 202610 min read

Governed Implementation

Design Without Governance Is Decoration

McKinsey says AI's scaling problem is a design problem. They're half right. Design is the interface layer of governance.

Mar 12, 20269 min read

Operating AI

The Operations Response: Amazon, Kubernetes, and the Week AI Hit Production Reality

Amazon mandates senior sign-off on AI code. Kubernetes builds AI governance into Gateway API. Code quality becomes ops infrastructure.

Mar 12, 202610 min read

The AI Control Problem

Half Your Benchmarks Are Wrong: What Happens When AI Measures Itself

METR finds 24pp gap between benchmark scores and real maintainer decisions. Anthropic quantifies 6pp infrastructure noise. PromptFoo joins OpenAI.

Mar 12, 20269 min read

Governed Implementation

Everyone Has 30% AI Code. Nobody Knows Who Governs It.

Uber and Stripe built governance into infrastructure. Microsoft hired a quality head. The pattern reveals what most companies are missing.

Mar 11, 202612 min read

The AI Control Problem

Builders, Reviewers, and the Governance Nobody Mentioned

Harrison Chase says coding agents split teams into builders and reviewers. The data shows a third role is missing: the one that decides what 'good' means.

Mar 11, 20269 min read

The AI Control Problem

AI Governance IS Cybersecurity

The org chart still separates them. The attackers don't. Why treating AI governance and cybersecurity as distinct functions creates structural vulnerability.

Mar 10, 202611 min read

Operating AI

The Agent Operations Paradox: More Agents, More Work, Less Uptime

Linear treats agents as team members. OpenAI can't hold three-nines. And AI creates more work, not less. Operations discipline is the missing piece.

Mar 9, 20269 min read

The AI Control Problem

From Issue Title to Malware: The Five-Step AI Supply Chain Attack

Clinejection turned a GitHub issue title into 4,000 compromised machines in five steps. Combined with Cloudflare's 2026 data, the pattern is clear.

Mar 9, 202610 min read

Governed Implementation

What 220 Controls Teach About Building AI Governance Frameworks

GitLab cut SOC controls by 58% with a custom framework. AI can now rewrite GPL code in days. Both stories reveal how governance actually works.

Mar 9, 20269 min read

The AI Control Problem

The Maintenance Crisis: Why AI Agents That Fix Bugs Create More Bugs

SWE-CI proves 75%+ of agent fixes introduce regressions over time. One-shot benchmarks hide the real problem: cumulative code decay.

Mar 9, 202611 min read

The AI Control Problem

The Mexican Standoff: What Happens When Everyone Can Do Everyone's Job

AI coverage hit 75% for programmers with zero unemployment increase. The threat is not job loss. It is role collapse without governance.

Mar 9, 20269 min read

The AI Control Problem

The $6-to-$1 Ratio: Why AI Is Turning Software Companies Into Services Firms

For every $1 spent on software, $6 goes to services. AI can deliver outcomes at software margins. The next trillion-dollar company already knows this.

Mar 9, 202611 min read

The AI Control Problem

The $500 Billion Question: Why Governance Is the Only AI Moat Left Standing

Hyperscalers spent $443B while 42% of companies abandoned AI initiatives. The surviving moat is not the model. It is governance.

Mar 5, 202610 min read

Engineering Notes

The Harness Difference: When 42% Becomes 78% Without Changing the Model

Same model, same benchmark. Claude Opus 4.5 scored 42% and 78% depending on harness. The model is not the product.

Mar 5, 202610 min read

Operating AI

From In-the-Loop to On-the-Loop: What Companies Running AI Agents Actually Do Differently

CircleCI data: fewer than 1 in 20 teams ship at AI speed. The ones that do engineer systems, not review diffs.

Mar 5, 202610 min read

The AI Control Problem

The Verification Escalation: Why the Window for AI Safety Infrastructure Is Closing

Faros.ai: 98% more PRs merged, 91% more review time. Leo de Moura says proofs must replace review. The IPO clock is ticking.

Mar 5, 202610 min read

The AI Control Problem

The Specification Problem Is the Governance Problem

AI-generated code can be mathematically proven correct. But correct according to what? The spec encodes values. That makes it governance.

Mar 4, 202610 min read

The AI Control Problem

You Are Not Killing Code Review. You Are Renaming Governance.

Aviator's CEO says code review is dead. His five-layer replacement is governance by another name.

Mar 4, 202612 min read

Engineering Notes

What Stripe's Agentic Layer Reveals About the Next Engineering Paradigm

Stripe's 1,300 weekly agent PRs prove the point: the architecture around the AI matters more than the AI itself.

Mar 4, 202612 min read

Operating AI

The 65% Rule: Why Production AI Is Becoming Mostly Code

Production AI systems converge toward hybrid architectures where deterministic code handles most work. The moat is not AI. It is governance.

Mar 3, 202610 min read

Operating AI

The Governance Loop Hidden in Your Agent Monitoring

Chase's production improvement loop is a governance framework in disguise. The convergence of observability and governance changes how you run AI.

Mar 3, 20269 min read

Governed Implementation

Agent Specs Are Governance Artifacts

Three independent frameworks converge on the same conclusion: agent specs are not documentation. They are auditable, enforceable governance infrastructure.

Mar 3, 202612 min read

The AI Control Problem

The $4 Deanonymization Problem and What It Reveals About AI Governance

LLMs can re-identify anonymous users for $4 per person. The real problem is not the capability. It is three governance failures converging at once.

Mar 3, 202611 min read

Engineering Notes

Seeing Like an Agent: What Claude Code's Tool Design Reveals About Building AI Systems

Anthropic's Claude Code team shares 4 tool design lessons. First-party evidence confirms: tools shape agent behavior more reliably than instructions.

Mar 3, 202610 min read

Governed Implementation

Your Agent's Personality Is a Governance Layer

OpenAI published four personality profiles for AI agents. They missed the point. Personality is fleet-wide behavioral governance, not prompt cosmetics.

Feb 27, 202610 min read

The AI Control Problem

AI Agents Have Opinions: What Claude Code's Tool Picks Reveal About Ungoverned Delegation

2,430 Claude responses reveal decisive tool preferences. GitHub Actions 94%, Express 0%. Training data is hidden policy shaping your architecture.

Feb 27, 202610 min read

The AI Control Problem

The Amplifier Effect: Why Your Org Chart Matters More Than Your AI Stack

AI doesn't create new organizational dynamics. It accelerates existing ones. The data reveals why governance is the input, not the output.

Feb 27, 202611 min read

The AI Control Problem

The AI Workforce Reckoning: When Irreversible Decisions Meet Unproven Capabilities

Block cut 40% of staff betting on AI. Oxford Economics says most AI layoffs are fiction. The governance gap between the two is where organizations fail.

Feb 27, 202611 min read

Operating AI

Automatic Rollback Is Necessary but Not Sufficient: The Missing Governance Layer for AI Deployments

AWS and New Relic ship automated rollback. But error-rate triggers cannot catch AI's hardest failure: plausible wrong answers that return HTTP 200.

Feb 27, 20269 min read

The AI Control Problem

When Enabling AI Changed the Rules: The Credential Escalation Pattern

Google API keys silently gained Gemini authentication. 2,863 keys found exposed. Enabling AI retroactively changes security assumptions.

Feb 27, 20269 min read

Governed Implementation

McKinsey Measured the Wrong Thing

Executives self-report 16-45% AI gains. Controlled trials show 19% slowdown. The perception mismatch is not noise. It is missing infrastructure.

Feb 27, 20268 min read

Operating AI

The Operations Tax: What Happens When AI Agents Hit Production Without Governance

MCP wastes 15,000 tokens per session. The fix removes the governance layer. This tension defines AI operations.

Feb 27, 20269 min read

The AI Control Problem

When AI Builds Its Own Tools: The Recursive Trust Problem

AI that writes its own code breaks the verification chain that made software trustworthy. The fix is governance, not more AI.

Feb 27, 20269 min read

Governed Implementation

Skills Are Not Replacing Agents. They Are Making Agents Governable.

The Agent Skills standard solves what monolithic agents never could: modular, auditable, version-controlled AI capabilities.

Feb 27, 20267 min read

Governed Implementation

Agent Frameworks Are Free Because You Are the Product

Hyperscalers give away agent SDKs to sell runtime. The real contest is governance: security, evaluation, context control. Bet on that layer.

Feb 26, 20267 min read

Governed Implementation

The Governance Layer That Was Already There

Three practitioners independently rediscovered the same truth: AI agents need engineering discipline, not new frameworks.

Feb 26, 20267 min read

The AI Control Problem

The Week Prompt Injection Became a Supply Chain Weapon

A GitHub issue title stole npm credentials and pushed malicious code to thousands. The attack surface is no longer the model.

Feb 26, 20268 min read

Governed Implementation

When the Spec Is the Product, Who Governs the Spec?

Spec-driven development compresses PM cycles. It also turns every unclear requirement into a production risk nobody reviews.

Feb 26, 20269 min read

The AI Control Problem

The Trust Gap Is the Governance Gap

84% of developers use AI tools. Only 33% trust the output. The gap is not about better tools. It is about missing governance.

Feb 26, 20268 min read

Governed Implementation

The Architecture of Agent Trust: Why Boundaries Beat Instructions

Reliable AI agents come from environmental constraints, not better prompts. Three independent sources converge on the same architectural principle.

Feb 25, 20269 min read

Operating AI

The Governance of AI Adoption: When Mandates Meet Mental Models

Tech giants are enforcing AI use through performance reviews. Mandates without cognitive alignment produce compliance, not capability.

Feb 25, 20268 min read

The AI Control Problem

When AI Measurement Breaks: METR's Dependency Problem

METR can no longer run controlled AI productivity experiments. Developers refuse to work without AI. This is a governance signal.

Feb 25, 20267 min read

The AI Control Problem

The Safety Paradox: How Competitive Pressure Dismantles AI Governance

Anthropic built its identity on AI safety. Now competitive pressure is forcing rollbacks. Voluntary commitments cannot survive market dynamics.

Feb 25, 20268 min read

The AI Control Problem

AI Doesn't Dilute Accountability. It Deletes It.

A builder spent $20K on AI credits in 3 months. The code shipped. What didn't ship: someone who wakes up at 3 AM when it breaks.

Feb 24, 20267 min read

The AI Control Problem

The Benchmark Is Contaminated. Now What?

OpenAI retired its own coding benchmark. 59% of tests were flawed, all frontier models contaminated. The measurement gap is a governance gap.

Feb 24, 20268 min read

The AI Control Problem

Cheap Code, Expensive Quality

Code generation dropped to near-free. Quality verification didn't. The gap between producing code and delivering good code is a governance problem.

Feb 24, 20267 min read

Operating AI

25 Hours, 13 Million Tokens: What a Codex Marathon Reveals About Agent Memory

GPT-5 Codex ran for 25 hours and generated 30K lines. The breakthrough wasn't the model — it was a 4-document memory system.

Feb 24, 20268 min read

Operating AI

40 Engineers, 1 PM: What OpenAI's Codex Team Reveals About AI-Native Organizations

OpenAI runs 40 engineers with 1 PM. The secret isn't talent density — it's hundreds of custom skills replacing coordination overhead.

Feb 24, 20269 min read

The AI Control Problem

Your AI Provider Is a Supply Chain Risk

Anthropic detected 24K fake accounts extracting Claude. If your competitive advantage runs on someone else's model, their security posture is yours.

Feb 24, 20269 min read

Governed Implementation

The Governance Outsourcing Problem: What OpenAI's Consulting Alliances Actually Mean

When consulting firms deploy your AI agents, they also define your governance. Enterprises need to decide who owns the rules.

Feb 24, 20268 min read

The AI Control Problem

The Organizational Debt of AI: Why 70% of Failures Have Nothing to Do with Technology

BCG found 70% of AI implementation hurdles are people and process. The real blockers are alignment gaps, dissolved boundaries, and broken talent pipelines.

Feb 24, 20269 min read

Operating AI

Your Agent Already Knows What's Wrong — You're Just Not Listening

Factory monitors 1,946 agent sessions daily and auto-resolves 73% of issues. The gap isn't AI capability — it's operational observability.

Feb 24, 20269 min read

The AI Control Problem

The $3 Trillion Stress Test Nobody Designed

Three AI IPOs will exceed a decade of US IPO capital. The financial system wasn't built for this transition speed.

Feb 24, 20268 min read

The AI Control Problem

The Governance Layer: What the Cybersecurity Selloff Reveals About AI-Era Security

Markets repriced $15B in cybersecurity value. The signal: detection is commodity. Governance is the moat.

Feb 23, 202610 min read

Operating AI

The Operations Discipline Gap: What Cloudflare and AWS Reveal About Governing Automated Systems

Two failures, one week. One was a code bug, the other an AI agent. Both reveal the same root cause: governance treated as afterthought.

Feb 23, 202611 min read

Governed Implementation

The Cage Pattern: What Agent Fleet Governance Actually Looks Like

Three companies running AI agents at scale converged on the same principle: maximum autonomy inside structural constraints.

Feb 23, 202610 min read

The AI Control Problem

Your AI System Needs an Enforcer Too

Design systems fail without active governance. AI systems fail the same way, for the same reasons. The enforcer pattern explains why.

Feb 23, 20269 min read

The AI Control Problem

The Distribution Follower: What a $20K AI Compiler Reveals About Governance

AI excels at reproducing known patterns. The governance question isn't whether AI can code — it's who decides what gets built.

Feb 23, 202610 min read

Engineering Notes

What 1.2 Million ChatGPT Responses Actually Reveal About LLM Citation Patterns

Separating verified signal from marketing noise in the emerging field of Generative Engine Optimization.

Feb 23, 202610 min read

Governed Implementation

Your Organization Is a Tactical Tornado

LLMs don't just write tactical code. They turn entire organizations into tactical tornadoes. The fix isn't better code review.

Feb 23, 20269 min read

The AI Control Problem

How AI Decides What to Quote — And Why It Matters for Governance

A study of 1.2M ChatGPT citations reveals predictable patterns. The governance question: if AI attention is an artifact, who governs the artifact?

Feb 20, 202610 min read

The AI Control Problem

Anthropic and the Pentagon: When AI Safety Becomes a Supply Chain Risk

The Pentagon may label Anthropic a supply chain risk over AI safety limits. Enterprise AI procurement now has a geopolitical dimension.

Feb 20, 202610 min read

Governed Implementation

Governance Is a UX Problem: What Apple's Agent Taxonomy Reveals

Apple mapped 55 UX features for AI agents. The finding most teams miss: governance that users cannot see is governance that does not work.

Feb 20, 20269 min read

Governed Implementation

The Containment Pattern: Four Approaches to Sandboxing AI Agents in Production

Cursor, Docker, Zenity, and Entire shipped four distinct containment layers in one week. The shift from approval fatigue to trust boundaries is here.

Feb 20, 202611 min read

The AI Control Problem

The Identity Problem: Why Developers Resist AI Tools and What It Actually Means

Developer AI resistance isn't Luddism. It's an identity crisis rooted in how craft communities process trust and truth.

Feb 20, 20269 min read

The AI Control Problem

When Interfaces Become Disposable: Where Does Governance Live When Users Skip Your UI?

Your UI was your last governance checkpoint. AI agents bypass it entirely. API governance is the new UI governance.

Feb 20, 20269 min read

Engineering Notes

The Architecture of Multi-Agent Systems: What Erlang, VS Code, and a Simpsons-Themed Loop Reveal

Erlang solved agent orchestration in 1986. Four converging signals show the industry is rediscovering supervision, isolation, and context discipline.

Feb 20, 202612 min read

The AI Control Problem

The Mythical Agent-Month: Why Your AI Strategy Needs Governance, Not More Tokens

Brooks's laws apply to agents. The brownfield barrier, the 1/9th problem, and 90% zero-ROI data show why governance beats parallelism.

Feb 20, 202612 min read

Operating AI

The Operations Gap: What Anthropic's Autonomy Study Reveals About Running AI at Scale

Anthropic studied millions of agent sessions. Experienced users grant 2x more autonomy. The real gap isn't trust — it's operations.

Feb 20, 202610 min read

Operating AI

Three Prices for One Agent: What Salesforce's AgentForce Pricing Reveals About AI Economics

When the biggest SaaS company on earth can't standardize agent pricing, your enterprise can't standardize agent cost governance.

Feb 20, 20269 min read

Operating AI

Your Agent Remembers Everything. Who Governs That?

Agent memory is the next governance frontier. Four architectures, four risk profiles — and nobody is auditing any of them.

Feb 18, 202610 min read

The AI Control Problem

The Governance Inflection Point: Sonnet 4.6 and the Cost of Near-Frontier Intelligence

When mid-tier models match flagships at one-fifth the cost, the governance question shifts from adoption to control velocity.

Feb 18, 20268 min read

The AI Control Problem

The Automation Curve Is Really a Governance Curve

McKinsey's 6-level framework shows what AI agents can do. It doesn't show how to choose or enforce the right level.

Feb 17, 202610 min read

Governed Implementation

The Rise of Agentic Platforms: Why Governance Is the Product

Bounded autonomy is the right design target for AI agents. Platform engineering is the governance layer most organizations already have.

Feb 16, 202610 min read

The AI Control Problem

When AI Builds Itself: The Governance Gap in AI-Assisted Development

Cognition uses Devin to build Devin. The real story isn't the recursion — it's the widening gap between code generation speed and review capacity.

Feb 16, 202610 min read

Engineering Notes

Your Training Data Is Going Stale. Composition-RL Shows What to Do About It.

Training data goes stale fast. An ICML 2026 paper shows how composing existing problems extends its life and why smaller models can win.

Feb 16, 202610 min read

Engineering Notes

Harness Engineering Is Not New — But Naming It Matters

OpenAI calls it harness engineering. Anthropic calls it effective harnesses. The discipline is old. The recognition is overdue.

Feb 16, 202610 min read

The AI Control Problem

Delegation Is Not Decomposition: What Google DeepMind Gets Right About AI Agents

DeepMind reframes multi-agent AI as a governance problem. The diagnosis is brilliant. The solutions are speculative.

Feb 16, 202610 min read

Governed Implementation

The Governance Gap in AI Output Quality: What Generic Design Reveals

Anthropic published a guide to fix generic AI designs. The real lesson: output quality requires the same governance as safety and compliance.

Feb 15, 20269 min read

The AI Control Problem

Cognitive Debt: The Invisible Cost of AI-Generated Code

AI code can be clean and still dangerous. When teams lose understanding of their own systems, governance is the only fix.

Feb 15, 202610 min read

The AI Control Problem

The Adolescence of Technology: What Amodei Gets Right, What He Misses, and What It Means for You

Dario Amodei warns about AI risks from the inside. His essay is essential reading — but enterprise leaders need more than policy frameworks.

Feb 14, 202612 min read

Engineering Notes

Agents Are Not Tools — But Sometimes They Should Be

The agent-vs-tool debate masks a deeper architectural question. The real skill is knowing when each pattern applies.

Feb 14, 202610 min read

The AI Control Problem

The AI Intensity Trap: When Productivity Gains Become Workload Creep

Berkeley researchers found AI intensifies work, not reduces it. The real finding isn't about AI — it's about governance.

Feb 14, 202610 min read

The AI Control Problem

What CEMEX's Financial AI Agent Actually Reveals About Enterprise Readiness

CEMEX built an AI agent for executives. The real story is what it exposes about governance gaps most companies ignore.

Feb 14, 20268 min read

Engineering Notes

Continuous AI: What GitHub's Agentic Workflows Actually Change

GitHub adds AI agents to Actions. The real shift is not markdown over YAML -- it is making non-deterministic work systematic.

Feb 14, 202610 min read

The AI Control Problem

Vertical AI and the Governance Gap in Professional Services

Vertical AI competes for personnel budgets, not IT budgets. That changes governance from a compliance exercise to an operational necessity.

Feb 14, 20269 min read

The AI Control Problem

Agency in the Age of AI: Why the Real Risk Isn't Intelligence

The tools that reward agency quietly erode it. Why AI governance must protect human decision-making, not just automate it.

Feb 13, 20268 min read

Operating AI

Three Models, One Week: Separating Signal from Noise in February 2026

OpenAI, Google, and Anthropic released frontier models the same week. Here is what actually matters for practitioners — and what is marketing.

Feb 13, 202614 min read

Operating AI

The AntFarm Pattern: What Specialized Agent Teams Get Right and What They Miss

AntFarm solves context degradation with agent specialization. But specialization without governance creates a different kind of fragility.

Feb 13, 202610 min read

The AI Control Problem

OpenClaw Is Not Claude Code — And That Confusion Tells You Everything

A viral article about AI governance confused two different projects. The error reveals how far the market is from understanding what it's trying to govern.

Feb 13, 202610 min read

The AI Control Problem

Your Product Team Was Designed for a World That No Longer Exists

Product teams face the biggest structural shift since Agile. The winners won't have the best AI. They'll have the best governance.

Feb 13, 202610 min read

The AI Control Problem

The Dead Companies Walking: What Steve Yegge Sees Coming

Yegge predicts 50% engineering cuts and eight levels of AI adoption. The real insight is about organizational absorption, not speed.

Feb 13, 202610 min read

The AI Control Problem

The AI Verification Debt

96% of engineers distrust AI output. Only 48% verify it. The gap is not a discipline problem. It is a governance failure.

Feb 12, 202613 min read

Operating AI

When AI Builds at the Speed of Thought, Who Decides What Gets Shipped?

60 agents, 77 overnight PRs, 33% rejected. Speed without governance is just expensive chaos.

Feb 12, 202610 min read

Governed Implementation

The Thinking Gap Is a Governance Gap

Every concern about AI code generation maps to a governance failure, not a technology deficiency. The question was never whether to use AI.

Feb 12, 202611 min read

Engineering Notes

WebMCP: Every Website Just Became a Tool for AI Agents

Chrome's new WebMCP standard lets websites expose structured tools to AI agents. What this means for the web.

Feb 12, 20268 min read

The AI Control Problem

Why Your AI Fails 40% of the Time

Benchmarks show sub-1% hallucination. Real-world tests show 40-60% failure. The gap is not about models. It is about process.

Feb 12, 202610 min read

The AI Control Problem

The Cloud Agent Thesis Is Half Right

Nader Dabit's four properties of cloud agents are real. They're also the four reasons you need governance before scale.

Feb 11, 202611 min read

The AI Control Problem

Don't Fire Your Team. Govern Your AI.

Claude Cowork is powerful. But it shipped with known vulnerabilities. Here's how to adopt AI workflows without losing control.

Feb 11, 202610 min read

Operating AI

The Most Governed Software Factory You've Never Heard Of

StrongDM says no human writes or reviews code. Look closer: every technique is governance in disguise.

Feb 11, 202610 min read

Governed Implementation

Your Style Guide Is a Governance Layer

Product teams encoding brand rules into CLAUDE.md are doing governance-as-code. Content ops proves the pattern.

Feb 11, 20269 min read

The AI Control Problem

Company as Code: The Missing Governance Layer for AI Agents

Why codifying your organizational structure matters more for AI agent governance than for compliance automation.

Feb 10, 20268 min read

The AI Control Problem

Your AI Stack Has a GitHub Actions Problem

Five companies exist just to make GitHub Actions faster. When workarounds become an industry, the problem is governance, not tooling.

Feb 10, 20268 min read

Engineering Notes

The February 5th Convergence: What GPT-5.3-Codex and Opus 4.6 Reveal

Both OpenAI and Anthropic released frontier models the same day. The real story is not which won — it is what their convergent bets tell practitioners.

Feb 10, 202612 min read

Operating AI

Agent Teams and the Shift from Writing Code to Directing Work

What a 100,000-line compiler built by 16 AI agents reveals about the future of software engineering and the governance it demands.

Feb 6, 202610 min read

Engineering Notes

Your Codebase Already Has an AI Governance Layer. You Just Don't Know It.

Stanford data shows codebase quality predicts AI productivity. The governance infrastructure was always there: linters, tests, type safety.

Feb 6, 20269 min read

The AI Control Problem

Agentic Engineering: Why What You Call AI-Assisted Development Matters

Osmani's agentic engineering framework reveals why naming your AI practice shapes governance, accountability, and results.

Feb 5, 202610 min read

The AI Control Problem

The AI Adoption Spectrum: What the 6x Productivity Gap Really Means for Enterprises

OpenAI data shows frontier workers are 6x more productive. The gap is real, but the binary framing is wrong.

Feb 5, 20269 min read

The AI Control Problem

The Benchmark Paradox: What AI Code Review Scores Actually Tell You

Every vendor wins their own benchmark. Academic tests show 3x lower scores. The gap reveals what enterprises need to govern.

Feb 5, 202611 min read

Engineering Notes

When AI Passed the Exam, The Exam Was the Problem

Anthropic retired a take-home exam after Claude matched top human performance. The real lesson isn't about AI speed. It's about governance.

Feb 4, 20267 min read

Engineering Notes

What Claude Code's /insights Reveals About Measuring AI-Assisted Development

A technical breakdown of the /insights command architecture and what its 6-stage pipeline teaches about observability in agentic workflows.

Feb 4, 20268 min read

Engineering Notes

What the Codex Agent Loop Reveals About Building Production AI Agents

An engineering teardown of OpenAI's open-source agent architecture and what it means for teams building their own.

Feb 4, 202610 min read

Engineering Notes

Context Is the New Perimeter: Why AI Agent Governance Starts in the Context Window

Tool descriptions steer agents. Context rot erodes control. The real governance challenge is not protocol choice — it is what the agent sees.

Feb 4, 202610 min read

Governed Implementation

The Developer's Real Edge Isn't Business. It's Governance.

AI is reshaping software careers. The advice to 'learn business' misses the point. The premium is in governance-aware architecture.

Feb 4, 20269 min read

The AI Control Problem

The Pinhole View of AI Value: Beyond Headcount Reduction

Kent Beck's NPV framework reveals why companies fixated on headcount cuts miss three out of four AI value levers.

Feb 4, 20269 min read

The AI Control Problem

The Phase Shift in Software Engineering

What Karpathy's 80/20 flip reveals about the gap between AI capability and real enterprise adoption.

Feb 2, 202612 min read

Governed Implementation

Product Discovery in the AI Era: When Building Is No Longer the Hard Part

AI compresses feasibility, viability, and usability risk. Desirability becomes the only differentiator. What changes for product teams.

Feb 2, 20269 min read

Governed Implementation

AI Agent Orchestration: From Prototype to Production

Why AI agents fail in production and how orchestration fixes it. Temporal, Conductor, and LangGraph compared.

Jan 29, 202612 min read

Governed Implementation

Agents That Learn: The Missing Layer in AI Systems

Most AI agents forget everything between sessions. Learn how runtime learning transforms agents from tools into teammates.

Jan 29, 202612 min read

Governed Implementation

AI in Design Workflows: What Atlassian, Meta, Tesco, and Faire Are Doing Differently

How leading companies integrate AI into design processes with templates, instruction files, plugins, and internal chatbots.

Jan 29, 202610 min read

The AI Control Problem

When Your AI Says Yes Too Often: The Hidden Risk of Digital Deference

Anthropic's research reveals AI can validate false beliefs, make moral judgments, and script personal decisions. Here's what leaders need to know.

Jan 29, 202610 min read

Engineering Notes

DeepSeek mHC: How a 1967 Technique Is Reinventing Neural Networks

DeepSeek's mHC innovation uses doubly stochastic matrices to stabilize deep networks. Technical analysis and strategic impact for leaders.

Jan 29, 202612 min read

Engineering Notes

Passive Context Wins: Why AGENTS.md Outperforms Skills in AI Agent Evals

Vercel study shows 100% pass rate with AGENTS.md vs 53% for skills. The problem: agents never invoked the skills they needed.

Jan 29, 202610 min read

Governed Implementation

Against Cleverness: Design Principles for AI in Complex Systems

Why conservative design beats clever automation. Lessons from nuclear engineering for building AI systems that fail gracefully.

Jan 28, 202610 min read

Engineering Notes

Claude + FigJam: When AI Outputs Stop Dying in Chat Logs

AI conversations generate insights that remain trapped in chat logs. FigJam integration creates shareable team artifacts.

Jan 28, 202610 min read

Governed Implementation

Memory for AI Agents: Benchmarks, Architectures, and a Surprising Discovery

A simple filesystem outperforms sophisticated memory solutions. Discover what benchmarks reveal about memory architectures for AI agents.

Jan 22, 202615 min read

Governed Implementation

AI Agents for Product Managers

How to use AI that acts and free up time for what matters. Use cases and tools for PMs without code.

Jan 22, 202610 min read

Engineering Notes

AI Code Review: Lessons from Cursor BugBot

How BugBot evolved from static pipeline to autonomous agent and what engineering teams can learn about AI-powered DevEx.

Jan 22, 202612 min read

The AI Control Problem

AI-First Culture in the Enterprise: What Ably Discovered

How Ably built an AI culture that works and why 70-85% of AI transformations fail. Practical lessons from a real case study.

Jan 22, 202612 min read

Governed Implementation

Claude's Constitution: Lessons for Enterprise AI Governance

Analysis of Anthropic's 23,000-word framework and how to apply its principles to corporate AI governance.

Jan 22, 202612 min read

Engineering Notes

Context Engineering for AI Agents: Lessons from Azure and Manus

How to manage context in production agentic systems. Seven lessons from Azure SRE Agent and advanced Manus techniques.

Jan 22, 202614 min read

Engineering Notes

FastMCP 3.0: The New Architecture for Production MCP Servers

FastMCP 3.0 rebuilds the framework around three fundamental primitives. Understand Components, Providers, and Transforms for enterprise.

Jan 22, 202614 min read

Governed Implementation

Genie: How Uber Organized Data for AI That Works

Insights from Uber's Gen AI on-call copilot. RAG vs fine-tuning, Spark pipeline, and the quality secret.

Jan 22, 202612 min read

The AI Control Problem

Growth Is Now a Trust Problem

Why traditional marketing channels are collapsing and how to build trust-based growth in the AI era.

Jan 22, 202610 min read

The AI Control Problem

Leadership in the AI Era: Why Human Skills Matter More Than Ever

AI can execute tasks at impressive speed, but it still cannot do the hard work of leadership. Discover the three exclusively human domains.

Jan 22, 202614 min read

Engineering Notes

MCP Design Patterns: Building Servers That Agents Can Actually Use

97 million monthly downloads. Also blamed for hallucinating agents. The problem is not the protocol - it is your server design.

Jan 22, 202612 min read

Engineering Notes

How to Build Self-Improving Coding Agents

If you repeat the same instructions every session, you are not developing with AI. Learn how to create systems that learn from each interaction.

Jan 22, 202612 min read

Governed Implementation

How to Write Specs for AI Agents

The definitive guide for specifications that work. 5 principles tested by Google and GitHub engineers.

Jan 22, 202612 min read

Engineering Notes

CLAUDE.md: The Instruction Manual for Your Code Assistant

Practical guide to creating an effective CLAUDE.md. Best practices, structure, examples, and common mistakes when setting up Claude Code.

Jan 19, 202612 min read

The AI Control Problem

Measuring AI in Software Development: What the Data Actually Shows

The productivity gains are real, but so is the perception gap. Here's what 600+ organizations reveal about AI measurement.

Jan 19, 202610 min read

Governed Implementation

Agentic AI: When Machines Don't Just Chat, But Act

The difference between AI that responds and AI that acts. How agentic systems transform expectations and deliver 4x productivity.

Jan 13, 20269 min read

The AI Control Problem

Agentic AI in Corporate Credit: 50% Faster Risk Analysis

Deutsche Bank case study: agentic AI cuts credit analysis time by 50% and boosts productivity 80%. See the multi-agent architecture.

Jan 13, 20269 min read

Engineering Notes

Claude Cowork: When AI Finally Works Alongside You

Anthropic releases Claude Cowork: agentic automation for everyone. Technical analysis, real use cases, and what actually works.

Jan 13, 20269 min read

Governed Implementation

Google UCP: Universal Commerce Protocol for Agentic AI Shopping

Google UCP enables AI agents to complete purchases conversationally. See how retailers integrate and what it means for e-commerce.

Jan 13, 20268 min read

Governed Implementation

The Anatomy of AI Agents: How Machines Sense, Think, and Act

From virtual assistants to autonomous vehicles, AI agents are everywhere. Learn the three-stage Sense-Think-Act architecture.

Dec 24, 20257 min read

The AI Control Problem

The Future Arrives Fast: 6 Radical Trends That Will Define 2026

Beyond language model hype, six interconnected forces — AI, geopolitics, economics, and demographics — converge to fundamentally transform our society.

Dec 22, 202510 min read

Governed Implementation

AI Agents: The New Frontier of the Job Market

With salaries reaching $4,000/month, AI agent specialists are among Brazil's most sought-after professionals. But what does this mean for your company?

Dec 22, 20256 min read

Operating AI

The 5Rs Framework: Why Most AI Initiatives Fail

AI projects fail due to organizational deficiencies, not technical ones. The 5Rs Framework transforms pilots into business results.

Dec 19, 20258 min read

Governed Implementation

One Year of Agentic AI: Six Lessons from Those Doing the Work

After a year of agentic AI projects, clear patterns emerge. Six fundamental lessons for capturing real value with autonomous agents.

Dec 19, 20258 min read

The AI Control Problem

AI Risk Management: A Framework for Companies That Cannot Afford to Fail

AI systems present unique risks. Learn the 7 characteristics of trustworthy AI and the risk management framework.

Dec 15, 20259 min read

Governed Implementation

From Pilot Purgatory to Production: A Practical Guide to AI Agents

Why two-thirds of organizations are stuck in pilot purgatory, and how to join the 8.6% that reach production with AI agents.

Jan 13, 202512 min read

The AI Control Problem

5 Counterintuitive Paradigms: How High-Performance Companies Lead

CTOs as growth architects. Governance as accelerator. Outcome roadmaps. 5 paradigms that separate high-performance companies from the rest.

Jan 13, 20259 min read

The AI Control Problem

The False Choice: Speed and Safety Were Never Enemies

The AI market tells you to choose between moving fast and staying safe. They're wrong. Here's why governance is architecture, not friction.

Dec 1, 20248 min read

By Topic

View all

Ready to assess your AI readiness?

Start with a four-week assessment. Clear answers about where you stand and what to do next.

Schedule Your Assessment Call