Eight weeks, 866 agent-written PRs: size predicted the reviewer's flags, the model mostly didn't
Oct 4 Zest (Ilya Volodarsky): every agent PR in two repos carries a self-written session block. Median 37 minutes from open to merge. Matched on repo, engineer and week, the authoring model barely moved the AI reviewer's should-fix rate, while 1,000-line PRs were flagged 36 points more often than sub-300-line ones.
Since August 12, every pull request a coding agent opens in Zest's two main repositories ends with a short block the agent writes itself: the harness, the model, minutes spent planning, coding and testing, the task type and the test tools it ran. Ilya Volodarsky's write-up covers the 866 of 1,570 merged PRs through October 4 that carry the block. It's one small team's sustained load over eight weeks, not a demo week, and that makes it useful.
The headline numbers:
- Merge time. The median agent PR merged 37 minutes after it was opened, for both Claude Code (822 PRs) and Codex (31). By model the medians ran from 25 to 47 minutes on PRs of similar size.
- The model changed under them twice in four weeks. Any cost or speed figure "per model" goes stale unless each session records which model it ran on.
- Coding was 43% of session time. Local testing took 29% and planning 22%; the median session behind one PR ran 48 minutes.
- Tests travel with fixes. 93% of the bug-fix PRs touched a test file, and 86% of PRs with a block did, against 61% of PRs merged in the same weeks without one.
Why it matters for code review
Every PR also gets a first pass from an AI reviewer. Raw should-fix rates ranged from 23% to 55% by authoring model, which looks like a model story until you control for when the work happened. Matched on repository, engineer and week, the Codex-versus-Claude Code gap shrank to one point with a wide interval, and only one model gap barely survived. Size did survive: PRs of 1,000 lines or more were flagged 36 points more often than PRs under 300 lines (interval +27 to +43), and the gradient held in August and late September alike. Under 100 lines, 11% were flagged; between 1,000 and 3,000, 57%.
The review loop after the first pass is mostly agent-to-agent. In the 30 days to October 4, 58% of reviewed PRs got at least one fix round; of 2,273 findings the agents answered, they fixed 80%, declined 18% and handed 1.6% to a human. Above 3,000 lines the flag rate stops rising partly because each reviewer lane reads a capped slice of the diff, which is a quiet warning that a large PR can look cleaner simply because less of it was read.
For a steady always-on setup, the takeaway is that the lever you control is the shape of the work, not the model choice: small PRs, a recorded session block, and review run on your own merged history in matching weeks. Volodarsky lists the caveats himself: one team, self-reported minutes, and 704 PRs without a block. Pair with measuring the queue in review minutes rather than agent count and a 562-PR stretch that merged in small, flagged pieces overnight.