In February we pushed straight to main and deployed by hand. In November, one person approved 182 pull requests in a week, almost all of them written by one AI and judged by another. Along the way each pull request got about twenty times slower to merge, and we think that was the right trade.
This is our own repo's history, measured: 2,838 pull requests, every label and comment on them, 6,244 CI runs and nine months of deploy records.
1. Push to main (Feb 2 – Apr 7). No CI and no protection. 2.0 merges a day.
2. CI, but nothing enforcing it (Apr 8 – Sep 22). GitHub Actions, self-hosted runners and an automatic deploy to dev arrived in a single commit. Then nothing about the process changed for 5.5 months and 1,612 pull requests. 9.6 merges a day.
3. A ruleset and a merge queue (Sep 23 – Oct 13). Main finally got rules, and a label from a human became the only way in. 18.7 merges a day.
4. The Verifier (Oct 14 – Nov 6). A second agent reads every pull request before the human does, and the queue refuses anything without its PASS. 16.7 merges a day.
5. Circuit (Nov 7 – today). Issues as the unit of work, a PM session that never writes code, one builder per slice, and GitHub as the only channel between agents. 21.7 merges a day.
Pull requests merged per month:
Median time from open to merge, by era:
Read those two lists together. Throughput and latency rose at the same time, and that only looks like a contradiction if a pull request is a person's afternoon. When agents write the code, a waiting pull request costs nothing but calendar time, and you can have 36 of them open at once.
For 5.5 months the process had one human-shaped hole in it. An agent would open a pull request. Another would merge first. Someone had to ask each open session whether its branch needed a rebase, wait for the answer, and then press the merge button. That someone was the only human on the team, working as a message bus.
The ruleset says a branch must be up to date with main before it merges. On its own that makes the relay worse, since every merge now invalidates every other open branch. The queue is what makes the rule livable: it takes approved pull requests one at a time, updates each, waits for CI, and squash-merges.
Both shipped within 48 hours of each other. Median label-to-merge in that era was 6.9 minutes, with no human in it.
Who wrote each month's merged pull requests:
It was never one handover. The two traded the lead three times before the roles settled. Codex wrote the merge queue. Claude wrote the Verifier. Today Claude builds and Codex judges.
By October we could merge 23 pull requests a day, and our own audit said only 28% of them were right the first time. Three in four defects we caught, we caught after the merge. The human's attention was going to rework.
So we moved the defect-finding to before the label. Since October 13 the Verifier has issued 1,652 verdicts on 639 pull requests.
More than half of all verdicts are a FAIL. Only about 4 in 10 pull requests pass on the first read. That is where the hours in the list above come from: it is the cost of finding out before your users do.
The number we would put on a slide is this one. First-read PASS rate by size of change:
A pull request that passes on its first verdict merges in 2.1 hours. One that needs four or more takes 15.1. Small changes are not a style preference. They are the difference between an afternoon and a day and a half.
Under Circuit, the life of an average pull request splits like this:
CI is not on the list. It runs in 4 to 6 minutes at the median and fails 2 to 7% of the time.
Three things we did not expect to find:
Method: everything here comes from GitHub's own records for one private repository, February 2 to November 17, 2026. Builder attribution uses branch names and commit trailers. Verdicts are parsed from pull request comments. November is a partial month and Circuit is 11 days old, so read its numbers as early.
Made by @qa-pagemcp