Parcha says a read-only shell took its cheap-model code reviewer from wrong 70% of the time to 7%

AgentRun is an open-source control plane that sits on top of an existing harness: Claude Code, Codex, OpenCode or pi. It adds durable runs, approvals, spend caps and a learning loop. With learning turned on, the agent gets one more turn at the end of each run to say what it would keep: a note, a skill or a workflow. AgentRun grades the proposals and commits the keepers to the agent's own git folder. Once a learned skill has served three runs, the next run is asked to draft that job as a workflow, and the draft is dry-run before it is kept. The README's own line is "A lot of agents don't need to be agents."

The numbers are in the code review case study. Parcha says its bill from Greptile, a code review product, went from $180 a month to $6,451 after Greptile moved in March to a dollar for every review past 50 and Parcha's own agents "took off". It built a replacement on a model it names particle-1. One parent agent sends out eight reviewers on a read-only checkout of the pull request, five quick and three deep, checks every finding against the code and posts one review. Parcha puts the cost at about 18 cents a review. It tuned on 70 merged PRs and scored on the 30 newest PRs in the AgentRun repository, which it had kept aside. On serious bugs its reviewer caught 5 of 8 and Greptile caught 1; the README gives the overall result as 1.6 times the real bugs Greptile caught.

Two of its lessons carry over to any reviewer. The first: "Reading only the diff, the cheap model was wrong 70% of the time. With read-only bash on the code, 7%." The second concerns scoring: a timeout that kept reviews from posting once made a good version look 15 points worse.

This is a vendor measuring its own product on its own repository. The answer key of 78 bugs, 8 of them serious, was built from findings graded by Opus 5.5, with no human grading described, and eight serious bugs is a small base. The write-up does not define "wrong". We read the README and the case study. We did not run AgentRun or open its tuning tutorial. The Hacker News post had 48 points and 5 comments about 2.8 hours after it went up.

Why it matters: Parcha's review bill rose 36 times in a year under per-review pricing, and this is its account of replacing the paid reviewer with a tuned cheap model. Two details are cheap to try on your own reviewer: give it read-only shell access to the checkout instead of the diff alone, and count the reviews that never posted when you score it. The comparison with Greptile rests on 30 PRs and an answer key graded by a model, so read the 5-of-8 figure as Parcha's claim.