440 factory PRs, 418 merged by a human click, and two bugs the green suite certified
Oct 6 Codecot factory sheet: a spec-driven agent loop with a rival-vendor reviewer opened 440 PRs over five months and ran a product for 19 days unattended. Its own defect list shows why a passing suite is not the same as a checked claim.
Source: The Factory, On One Page
Volodymyr Pasichnyk publishes the full ledger of a spec-driven agent factory: a human writes a spec and clicks merge, agents plan, implement and run the tests, and a reviewer model from a different vendor reads the diff and posts findings before fix rounds land. Between 22 April and 2 October 2026 it turned 794 specs into 440 pull requests, of which 420 merged, 418 by a person clicking the button. It ran 1,686 agent rounds at roughly $2.49 a round, and 80% of 237 branches were green on their first CI run.
One worked pass shows the shape of steady throughput: spec merged at 12:43, PR open at 12:53, 13 review findings at 12:56, all answered by 13:06, a second review with 12 more findings and a "ready" verdict at 13:13, human merge at 16:41. Spec to PR took ten minutes; the person spent minutes of the four hours. Across the whole run, though, the median from merged spec to merged implementation was 60 hours: the human click, not the agents, sets the pace. Reviews were busy, with 3,037 findings in 705 reviews, 410 fix rounds, and 35 rounds refused by a guard for claiming "already satisfied" without changing anything. The same loop published a daily product for 19 days with nobody pressing anything.
The page is most useful for what it admits. Of 127 failed developer rounds, 51% were infrastructure (wrong commit, branch races, dirty checkouts) and none failed on an ambiguous spec. Two defects reached production because tests certified them: a zero-length digest window reported a quiet sky for 35 mornings, and a failed voice gate was treated as consent, so an eclipse that did not happen was published in four languages and found six days later by a viewer.
Why it matters for code review
This is what an always-on factory looks like when someone counts honestly: modest per-round cost, high first-pass CI, a cross-vendor reviewer doing most of the reading, and a human merge that becomes the queue. It also shows where continuous verification breaks. A green suite looks the same whether it tests the behavior or its own mistake, the same failure as the overnight PR whose eleven green tests never ran and 37 agents passing a suite that missed two bugs. The fix here is to make gates recompute figures from source and to treat a failed step as a stop, not as silence.