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

Task throughput per developer rose 33.7 percent in two years. Epics completed per developer rose 66.2 percent. Tasks with an associated pull request grew 210 percent at the team level. Over the same window, in the same organizations, deployments per week fell 11.7 percent and the lead time from code commit to production grew 480.4 percent. Every number comes from one dataset: Faros AI’s AI Engineering Report 2026, built on telemetry from 22,000 developers across more than 4,000 teams. The system produces more than it ever has and releases it less often than it did before AI arrived.

We wrote about this report’s quality story in June. The acceleration whiplash piece traced what AI volume does to review load and verification work, and it remains the companion to this one. This piece follows the other half of the telemetry: what happens to work between “started” and “running in production” when every developer feeds the pipeline a third more of it.

Easy to begin and hard to finish

The average task now spends 225.2 percent more time in progress. Time in wait states is up 81.8 percent. In-progress tasks with no activity for seven or more days grew 26 percent. Developers touch 67.4 percent more PR contexts per day, and work restarts are up 13.8 percent. Faros compresses the pattern into one phrase: work has become easy to begin and hard to finish.

Every item on that list sits downstream of code generation, in stages that run on human attention. Code review, QA, and the judgment call on whether a change is safe to release each have a fixed daily capacity, and each is now fed by a source that grew by a third. Queueing arithmetic does not negotiate. When the arrival rate rises against fixed service capacity, queues lengthen faster than linearly, waiting swallows the schedule, and throughput at the end of the pipe can fall even while input surges at the top. That is what a deployment counter reading minus 11.7 percent next to a task counter reading plus 33.7 percent looks like from inside a queue.

Starting has also become cheaper than finishing, which skews behavior. An agent opens the next task in seconds. Closing the previous one still requires a human to review the diff, exercise the judgment, and own the consequence. When starting costs nothing and finishing costs attention, the ratio of open to done drifts exactly the way the stall numbers say it drifted. The same scarcity explains why the organizations absorbing the most AI keep hiring: attention is the input that ran out, a pattern visible in Ramp’s headcount data on heavy AI adopters.

One caveat belongs in the open. Deployment frequency and lead time are measured by roughly 10 percent of the Faros dataset, the subset that instruments deployments. Within that subset the deployment drop is statistically significant, and Faros presents both figures as directional. Read them conservatively and they still point the same way as the wait-state and stall metrics, which come from the full dataset.

Maturity does not protect

DORA’s 2025 report, a survey of about 5,000 technology professionals, concluded that “AI’s primary role is as an amplifier, magnifying an organization’s existing strengths and weaknesses.” The comforting corollary spread fast: build strong foundations and AI makes you stronger. Faros’s telemetry lands directly on that claim and breaks it: “The organizations with strong pre-AI engineering practices see the same quality degradation as those without.”

The split is methodological. DORA measures what practitioners report perceiving; Faros measures what the pipeline did. Both can be honest at once. Amplification may hold for the outcomes people can feel, while flow degradation accumulates below the threshold of perception, one lengthening queue at a time. DORA’s own earlier work points the same direction: its 2024 report found that a 25 percent increase in AI adoption was associated with a 7.2 percent decrease in delivery stability. The telemetry extends a signal DORA surfaced first.

Why would mature organizations degrade like immature ones? Because maturity is calibration. A high DORA score certifies a pipeline tuned to a specific arrival rate of change: so many PRs per week, so much review capacity, gates sized for the flow they were designed under. AI moved the arrival rate and left the calibration in place. A review culture that is excellent at ten PRs per week is a queue at thirty. The stage gates that made the pipeline trustworthy are the same ones now metering it. Well-designed constraints do not resize themselves.

The redesign is a flow problem, and it has a price

Protection comes from redesigning the flow for the new arrival rate, and the redesign starts with instrumentation almost nobody has. Adoption dashboards measure the generation side: suggestions accepted, PRs merged, cycle time at the coding stage. Flow collapse shows up only in queue metrics: time in wait states per stage, count of in-progress items idle for seven days or more, deployments per week, commit-to-production lead time. If those four numbers are missing from the dashboard, the collapse stays invisible until customers report it. Due disclosure: Faros sells engineering-intelligence tooling, so its conclusion that you need flow visibility carries a vendor incentive; the deltas above are telemetry and stand regardless of whose dashboard you buy.

With the queues visible, the moves are unglamorous. Cap work in progress so that finishing outranks starting. Move verification capacity to where the queues actually form rather than where the org chart placed it. Automate the reversible parts of the release decision so human judgment is spent only on the irreversible ones. Each of those moves is a governance decision with a price attached, and pricing it explicitly beats absorbing it silently, the same discipline we argued for in the model threshold decision. The production side of this same volume story, where unfamiliar code starts paging humans who never wrote it, is mapped in the code tsunami piece.

Do this now

Pull two curves for the last four quarters. Curve one: any generation-side metric you already track, merged PRs or completed tasks per developer. Curve two: deployments per week and commit-to-production lead time. If curve one rises while curve two is flat or falling, flow collapse is already in progress, whatever your DORA score says. Then instrument the wait: time in wait states per pipeline stage, plus a weekly count of in-progress items idle for seven days or more. Set a work-in-progress cap wherever the wait concentrates, and recheck quarterly, because the arrival rate will move again with the next model generation.

Mature teams have earned the habit of trusting their own pipeline. The Faros data says that habit is now the risk. The pipeline was tuned for a volume that no longer exists.

---

This analysis synthesizes AI Engineering Report 2026: The Acceleration Whiplash (Faros AI, April 2026), State of AI-assisted Software Development (DORA / Google Cloud, September 2025), Accelerate State of DevOps Report 2024 (DORA / Google Cloud, October 2024).

Victorino Group helps engineering organizations redesign delivery flow for AI-scale code volume, from wait-state instrumentation to governed release paths. Let’s talk.