Four agents, 90 PRs a day, 15 merges: tier by risk surface and let revert rate move the line
Aug 10 steezr (Johnny Unar): an inherited client repo had 300+ open agent PRs and two exhausted senior reviewers. Path-based risk tiers took about 60% of agent PRs off human review; commit trailers recording provenance showed a 1.8% revert rate on auto-merged low-risk PRs against a 2.3% human baseline, and 11% on high-risk surfaces.
Source: The Review Queue Is Now Your Deployment Bottleneck
Johnny Unar's consultancy inherited a client codebase where four agents opened about 90 PRs a day, merges had flattened at roughly 15 a day, and the open count had passed 300. CI runners were saturated testing branches nobody would touch for a week, and the two senior reviewers had "quietly stopped sleeping" until someone rage-approved a batch to make the number drop. The review process was still the 2023 one: the same template, two required approvals, and whoever was free picked up the next PR.
What they changed
- Tier by what the diff touches, not who wrote it. A labeling Action reads a paths-to-tier YAML and marks each PR
risk:low,risk:mediumorrisk:high. Branch protection keys off the label. Low-risk surfaces auto-merge on green CI under a diff-size ceiling, medium gets one reviewer with a ten-minute target, and auth, payments and migrations always get a senior reviewer. About 60% of agent PRs stopped needing a human at all. - Record provenance at commit time. The agent wrapper adds git trailers for model, task id and human, agent or agent-assisted. That makes revert and defect rates sliceable by origin and tier.
- Make CI cheap for work that won't land. Run lint and type checks before the expensive suite, skip reruns when relevant paths haven't changed, cap concurrency per agent so human PRs aren't starved, and auto-close agent PRs with no activity after 72 hours.
Why it matters for code review
The provenance data is what moved the trust boundary. Auto-merged low-risk agent PRs were reverted 1.8% of the time, below the 2.3% human baseline on the same surfaces, so the auto-merge band widened. On high-risk surfaces the same model's revert rate was 11%, so a human stayed in the loop there. Unar's rollout advice fits steady always-on work: auto-merge one well-tested surface, watch its revert rate for a month, and widen only if it holds at or below the human rate. The numbers come from a consultancy's own client work, but the method is one you can run on your own merged history. Compare vendor revert rates across 37,000 agent PRs and four tiers of trust for verification.