One person runs a software factory: agents plan, implement, review and open the PR. The human sits on approvals and steering, not in the loop. The same discipline as everywhere on this shelf: nothing merges on an agent's word alone, and every number below is counted, not claimed.
Real stage names from the running system, with no invented screenshots. Press play to walk one spec through it.
Why a rival vendor in seat 04: measured on this factory's own production data, one model family reviewing its own code raised a blocking concern in 0% of rounds; a rival vendor on the same code: 43%. Review that cannot disagree is not review.
Counted from records on 2026-10-02, over 2026-04-22 to 2026-10-02. Each card names its own source and its own date, and the page source keeps the method behind each, so these rot loudly, not quietly. The whole sheet, on one page →
794 spec documents. A spec is its own pull request; the implementation is a separate one — acceptance criteria are never an afterthought.
440 pull requests opened by the factory, 420 merged, and 418 of those merged by a human clicking the button.
80% of factory branches were green on the first CI run, 90% in the main repository.
3,037 findings from the rival-vendor reviewer in 705 reviews, and 410 fix rounds that followed them.
Median 60 hours from a merged spec to the merged code that implements it. 58 specs shipped inside a day.
Same model family reviewing its own code: 0% blocking verdicts. A rival vendor on the same code: 43%. So the reviewer seat is always a rival.
127 of 882 developer rounds failed. Half of them were infrastructure. None failed on an ambiguous spec.
~140,000 lines of TypeScript, one person. Planned, implemented, reviewed and merged through the loop above.
The point of the factory is not to remove the human but to move them to the two seats where a human is irreplaceable.
Nothing merges without a click, and the click is informed: the PR arrives with the spec it implements, the reviewer's verdicts, and CI state attached. The same shape as this lab's publishing gate: verified first, approved second, shipped third.
Review findings are banked, not re-litigated: a concern raised once binds the follow-up work until it is fixed or explicitly waived. The review loop debounces to the branch head so agents converge instead of circling. The human writes specs and priorities; the loop turns them into merged code.
The factory itself is not exposed publicly: it holds write access to real repositories, and autonomy is earned, even here. What you can touch is this shelf: the same verify-then-approve discipline, live, in the cockpit. The factory's fuller story lives at codecot.com.