53 merges while he slept, and the costliest failure was a flaky test
Jacob Archer tears down one eight-hour night in which agent lanes merged 53 PRs with no human review. A batch bisector caught a real bug for free; a timeout on a busy machine stopped every merge for half an hour.
Source: 53 PRs While I Slept: A Teardown of My Best Overnight Run
Jacob Archer, a solo founder building Cities In My Pocket, went to bed on October 19 and woke to 53 merged pull requests: 43 in his main repo and 10 in a second one, all between 11 PM and 7 AM Eastern. He wrote and reviewed none of them. His post is a teardown rather than a brag, and every figure comes from GitHub or his own logs.
The setup has one assistant writing tasks into per-area queues, a supervisor handing work to idle lanes, coding agents opening PRs, an automatic reviewer, and a merge driver on his own machine. The supervisor launched 33 agent runs. Most PRs were one slice of a feature shipped behind an off flag, plus test-only PRs, two FAQs, and five fixes from a nightly bug sweep.
A PR has to pass three gates. The reviewer approves a specific commit, and a new push voids that approval, which happened 7 times after midnight. The merge driver applies ready PRs as a batch on top of main and runs the full check, splitting a failed batch in half until it finds the culprit. After a handful of merges it checks main itself, and a red main pauses everything.
The night tested both failure modes. Bisection isolated one residents PR that failed even alone; the system queued a fix to the lane that owned it, and it merged at 4:27 AM in a clean batch. That cost nothing. Then at 1:29 AM main went red because a single test timed out at 5 seconds on a loaded machine. Run alone it passed in 4.2. Nothing merged for about half an hour, which explains most of the dip from 4 merges at 1 AM to 2 each at 2 and 3 AM. Conflicts were routine: PRs that no longer applied cleanly were held and sent back, and seven came back fixed before 7 AM.
Why it matters for code review
This is what steady overnight throughput looks like when it works, and the lesson is that the volume rests on dull mechanics: small flagged PRs, approvals tied to a commit, batch bisection, and pausing on red. Archer's own ranking is the useful part. A real bug was cheap; a test that couldn't say whether main was healthy was the most expensive thing all night. Under sustained load, flaky checks stop being an annoyance and become the limit on throughput.
It is also a case with no human reviewer before merge. Archer's morning job was deciding which flags to turn on, so the off flag is doing the work a reviewer would do: nothing reaches a user until he chooses. Teams that can't ship dark have to put that judgment back before the merge.