Traycer's factory opened 332 of the 675 PRs a four-person team merged in 11 days, and the step that compounds is the factory fixing its own jobs
Tanveer Gill's build log: chores first, a blind reviewer that never sees the author agent's reasoning, the agent that opens a PR owns it to merge, a 4:24am user report merged in 23 minutes, and 43 of 332 factory PRs improving the factory itself.
Tanveer Gill of Traycer writes up the first eleven days of the team's software factory, which they define as a system where work enters on its own triggers, agents carry it all the way to proof, and people make the decisions. Over that stretch the factory opened 332 of the 675 PRs the four-person engineering team merged. 92% of factory PRs merged (28 closed unmerged), the median time from opened to merged was 1.6 hours, and the fastest user report reached a merged fix in 23 minutes: a Spanish-locale Windows install failure reported at 4:24am, matched to 200+ Sentry errors, with a failing test and a fix up in 8 minutes. The PR said plainly that nobody had a Windows machine for a full end-to-end run.
The build order is deliberately small. They started with chores where a wrong answer costs nothing to ignore (Sentry errors and flaky tests on day one), sandboxed every agent with read-only tokens, and routed every outbound message through a gateway with plain rules, daily budgets, and quiet hours. An LLM filter they tried first silently held 40 messages the jobs had been told to send, including real approval requests. Review runs in layers: several specialist reviewer agents, one blind reviewer that reads the diff and the original problem but never the author agent's reasoning, and people who read the proof. The agent that opens a PR owns it to merge, with a 15-minute sweep nudging anything stuck. 43 of the first 332 factory PRs were the factory improving its own jobs, and every job audits itself weekly.
Why it matters for code review
This is the case for steady, harnessed throughput over heroic sessions, with numbers: half of a small team's merged work came from a loop that runs while they sleep, and the compounding comes from the self-improvement step rather than from any one clever prompt. The review lessons travel well. Keep the reviewer blind to the author's argument so it can't be talked into agreeing, make the evidence faster to review than the diff, and treat "no action needed" messages as a defect: one job drafted 45 messages in three days and 36 said nothing. People keep product calls, auth, billing, releases, and anything irreversible. Read beside Pocock on the PR bottleneck and reviewer retros and Robson's gardener that samples merged work.