Faros's Speed Trap telemetry: per-PR incident risk settled, but PRs merged with no review are up 76.3% and QA time is up 300%

Faros's new report looks at teams that are already deep into AI adoption, using the most recent 12 months of telemetry from 22,000 developers across 4,000 teams. The headline is that the pipeline is moving but leaking. Faros describes it as unblocked, but leaky: more work is flowing through, but sacrifices are being made along the way to keep it moving.

Review is what gets skipped

Six months ago Faros reported a 31.3% rise in PRs merged without review; that figure has increased to 76.3%. The changes are also bigger, with average PR size up 71.8%. Faros's reading is that review capacity is not growing at the same rate, so under long queues and delivery pressure teams may reduce or skip review altogether to keep work moving, and those one-off decisions can become a default operating practice over time. The cost then shows up later in the line: time in QA is up 300.6%, because QA inherits the verification burden, where it is often slower and more expensive.

Safer changes, many more of them

Per-change risk has calmed down. In April incidents per PR were up 242.7%; now they are up just 14.5%. But monthly incidents are up 125.4%, because there are far more changes moving through the system. Faros's Vitaly Gordon argues the most expensive number is a different one: work restarts per developer, tasks sent back after reaching review or QA, are up 66.7%, against only 13.8% in April. A restart is paid for twice, once in the tokens for the wrong output and again for the fix, plus the human time to find the error.

Where agentic review fits

Many companies now run AI review on 50–80% of PRs, while agents open 13–14% of PRs at the leading edge. Teams that use agentic review heavily see faster first reviews and lower change failure rates, though Faros stresses the findings are correlational, and unreviewed merges continue to rise even where adoption is high. Its own recommendation still points upstream: the greatest leverage lies at the authoring stage.

For anyone running agents around the clock, this is the volume side of the verification problem. Steady output only compounds if the review and QA stages scale with it; otherwise the extra throughput turns into skipped review and restarted work. Compare Faros's earlier measurement framing for software factories.