Every number comes from our own, day-to-day delivery — the VaultPAM security product, built in the AI SDLC by a team where most roles are filled by AI agents. Sources are given next to each number; methodology at the bottom of the page.
A 0.6 h median from PR to merge does not mean "nobody is looking". It means machines are looking — immediately: verdicts, CI and gates run in minutes instead of waiting in a queue for someone's calendar.
68 QA stops are not an embarrassing statistic — they are a working brake. Each one is a defect that never reached the main branch, has a named reason and a ticket, and the work went back for a fix along the same auditable path. See such a return step by step in the walkthrough demo.
Our metrics come from a single team — so we set them against independent industry research. Its conclusions are not flattering for AI, and in our view rightly so: they show what happens when speed has no process. In March 2026 DORA added the other half — small batches, automated testing and fast feedback loops as the counterweight to destabilization. Every number below can be checked at the source — the links sit under each card.
AI adoption now correlates positively with delivery throughput and still negatively with delivery stability. 90% of respondents use AI at work, over 80% report a productivity gain — and 30% have little or no trust in AI-generated code.
Why we cite it: this is a change from DORA 2024, where throughput was negative too. The gap did not close, it moved: speed went up, stability did not follow. Gates address exactly that second half.
AI-co-authored PRs: 10.8 issues per PR against 6.5 in hand-written ones — roughly 1.7×. At the 90th percentile AI PRs reach 26 issues per change, more than double the human baseline.
Why we cite it: agent code needs harder verification, not softer. Our 68 QA stops (~1.2% of tasks) are not an embarrassment — they are the mechanism that neutralizes these statistics before merge.
Source: coderabbit.ai → State of AI vs Human Code Generation
Maintainer merge decisions ran on average 24 percentage points below the automated SWE-bench grader score. Control: known-good, human-written patches passed the grader 100% of the time and won reviewer approval 68% of the time.
Why we cite it: green tests are not a merge decision — for anyone, not even for a human. For agents the gap is ~24 points wider. That is why the gate verdict here is issued by a role other than the change's author.
Source: metr.org → Many SWE-bench-Passing PRs Would Not Be Merged