PR-to-merge is the factory bottleneck — specialize the loop
Aug 13 Augment: generation is cheap; review/verify/repair/decision are the drag. Cosmos Experts (Risk, Deep, Pair, Verifier, Fixer) keep working until a human can merge with evidence — loop engineering, not a better comment bot.
Source: The software factory needs a faster review loop: further optimizing the path from PR to merge
Akshay Utture’s Aug 13 Augment post starts where the tortoise thesis’s merge-side half lives: once agents write code, review, verification, repair, and human decision-making become the binding constraint. A standalone AI review bot that posts comments is useful and still a one-step tool. The drag is the chain of human handoffs after the diff exists — reconstruct intent, triage low-risk changes, apply feedback, fix CI and conflicts, exercise behavior end to end, gather evidence, re-review after each push. Their earlier Cosmos review system already claimed ~3× code output with shorter median merge time; this piece expands the same idea into a full PR-to-merge loop.
The architecture is specialization, not a smarter generalist. Risk Analyzer classifies and applies auto-approval policy. Deep Reviewer does exhaustive objective-correctness analysis against repo guidance (AGENTS.md / CLAUDE.md). Pair Reviewer reconstructs intent and flags judgment calls with a human. New pieces: Verifier exercises the change in a running environment and returns screenshots, logs, traces; PR Fixer owns mechanical follow-up (findings, CI, conflicts) from any PR, human- or agent-authored; a Review Dashboard makes fleet state legible; cosmos approve evaluates a configurable policy (ownership, current-head review clear, no contradictory runtime evidence) while a human still clicks Merge. None of the Experts merge. Tokens that cut human time in the bottleneck are treated as cost-per-successful-outcome spend, not waste to minimize.
Why it matters for code review
Burst Claude mastery still dumps unfinished work into the same PR form. Steady harnessed throughput only compounds if the merge path keeps working overnight and across volume — repair loops, runtime evidence, and policy checks that do not wait for a hero session. FOMO when agents aren’t working on your behalf is real; the counterfeit version is FOMO for more generation while the review loop stays serial and human-only. Measure whether your factory owns the path to a verified, understood, merge-ready change — not time-to-first-comment on an unread flood.