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

On 18 July we argued that the line between what an AI decides and what a human decides should be set once per action class, by blast radius and reversibility, and enforced in the harness instead of relitigated in a meeting. Three days later, Hud published four agentic workflow recipes as an MIT-licensed public repository. The first one is called Blast Radius. It scores a pull request’s production blast radius from 0 to 100.

We had nothing to do with it. That is what makes it useful. An argument that only its authors find compelling is a position. An argument that shows up independently as working source code is a design converging from two directions.

Hud is an early-stage startup with $21M raised, selling the exact capability these recipes showcase. Take that seriously as a bias, then read the recipes anyway, because they are MIT-licensed and short enough to audit in an afternoon. You can read the scoring logic, disagree with a weight, and change it. That is a different kind of artifact from a governance claim in a vendor deck.

Four recipes, and what each one decides

The repository ships Blast Radius, Weekly Report, Dead-Code Cleanup, and Rollback Check. Each is a named agentic workflow rather than a general assistant prompt, which already narrows what can go wrong.

Blast Radius takes a pull request and produces a 0-to-100 production impact score. Dead-Code Cleanup finds functions with zero invocations and opens a pull request removing them, which only works if something is counting invocations in production. Rollback Check evaluates a deployment and returns a verdict.

The counting comes from Hud’s product, which its own site names the Runtime Code Sensor: “a new runtime layer that runs with your code in production. It detects errors, performance degradations, and CPU spikes, capturing deep forensic context needed to agentically generate safe, code-level fixes.” The recipes sit on top of that telemetry. They are, in the plainest reading, a demonstration of why you would buy the sensor.

A number where we left a principle

Our post gave two axes and said to attach them to an action class in code, once. It did not say what attaching looks like. A 0-to-100 score is one answer, and it carries properties an argument does not. It is comparable across pull requests. It can be thresholded. A change to the threshold is a diff with an author and a date.

That last property is where this connects to the piece we published this morning. The per-action permission prompt is not a control, because it arrives at the moment of action, aimed at whoever happens to be at the keyboard, under whatever time pressure that person is under. A score computed before merge and compared against a threshold set weeks earlier is decided in advance, applies to a class rather than an instance, and leaves a record of who set it. Same governance intent, opposite failure profile.

The weights are also where the disagreement should happen. A scoring function that treats a migration file and a copy change identically is wrong in a way you can point at. Reading Hud’s weights and rejecting half of them is a more productive hour than debating whether blast radius matters.

Five verdicts in place of a boolean

We argued in February that automated rollback is necessary but not sufficient, because an error-rate trigger tells you something moved without telling you why, and binary revert is often the wrong response to a real signal.

Rollback Check returns one of five values: ROLLBACK, INVESTIGATE_OUTBOUND, INVESTIGATE_ENVIRONMENTAL, WARN, CLEAN. Two of those five say investigate and name the direction to look. INVESTIGATE_ENVIRONMENTAL covers the case where the deploy is innocent and the host, the dependency, or the traffic shape changed underneath it. WARN covers degradation that is real and below the threshold where anyone would revert.

Each verdict routes to a different human action. That is the graduated response our February post asked for, expressed as an enum instead of an essay. An enum is also testable. You can write a case that asserts a known environmental incident produces INVESTIGATE_ENVIRONMENTAL, and watch it fail when someone loosens the logic.

The gains land on a desk, the costs land in the system

DORA’s 2025 research, restated in its 2026 ROI report, ranks ten outcomes by effect size. Individual effectiveness is first. Software delivery instability is second. Throughput is seventh. In DORA’s words: “The largest effect was measured on individual effectiveness. The effect on instability was second largest, larger than the impacts on organizational performance, code quality, and so forth.”

Throughput did improve. DORA’s 2025 executive summary is explicit that “AI adoption now improves software delivery throughput, a key shift from last year. However, it still increases delivery instability. This suggests that while teams are adapting for speed, their underlying systems have not yet evolved to safely manage AI-accelerated development.” The survey behind that statement ran from 13 June to 21 July 2025 with 4,867 respondents across more than 100 countries, reported as standardized beta weights with 89% credible intervals rather than headline percentages.

Read the ranking as an accounting statement. The individual gets the largest benefit. The system absorbs the second largest cost. A blast radius score and an enumerated rollback verdict are both instability controls, which puts them exactly where the second-largest effect lives.

Where the recipes stop

May Walter, Hud’s co-founder and CTO, gave a talk in July describing why this is hard. Coding agents reason over code at the function and file level. Production metrics live at the service and endpoint level. The two do not speak the same language, and she calls the bridge between them prod-to-code. That framing is hers from the stage, not a documented product feature.

She names two failure modes worth stealing regardless of what you think of the product. The plausible unverified is a fix that sounds right and looks real and does not hold when checked against production. The lazy fix is an agent catching an exception rather than finding why it was thrown. Both survive code review by a tired human.

Her team deliberately does not auto-open pull requests: “no one wants to wake up for, you know, a rain of 80 pull requests, as small as small as they can be. That’s just not how people operate.” And on when autonomy is earned: “If something works 80% of the time and you’re using it with your cursor in your IDE, that’s fine because you’re there, you’re in context, and you can help fix and steer. If we’re talking about an automation that runs autonomously, we have to have a very high confidence.”

That threshold is the same one we drew by blast radius, arrived at from the operator’s side. Hud publishes a ZoomInfo case study claiming dependency lookup invocations dropped 98% and peak memory fell 62%, which is the vendor’s own account of a customer result rather than independent measurement, and does not change the threshold argument either way.

Do this now

Clone the repository and read the Blast Radius scoring rules against the action-class inventory you already have. If you do not have that inventory, the scoring rules are a serviceable first draft of one. Mark every weight you would change and why. The output of that hour is your own scoring function, which you now own.

Then do the smaller and more valuable thing. Find the place in your deployment pipeline where a boolean decides revert or proceed, and replace it with at least four named verdicts, each routed to a named owner. You do not need Hud’s runtime sensor to do that. You need to admit that “something is wrong” and “the deploy is wrong” are different findings that your pipeline currently collapses into one.

We wrote that the operational line is buildable today. Someone built a version of it and gave it away under a license that lets you audit the reasoning. The remaining work is deciding whether their numbers are your numbers.

---

This analysis synthesizes hud-agentic-workflows-recipes (Hud, July 2026), From Blind Spots to Merged PRs (May Walter, Hud, at AI Engineer, July 2026), The ROI of AI-assisted Software Development (DORA, Google Cloud, 2026).

Victorino Group helps engineering teams turn blast radius from a principle into a scored, thresholded control their pipeline enforces. Let’s talk.