Claude Code team hands most code reading to Claude

Anthropic’s Claude Code team now has Claude do most of its code reading and keeps human approval for changes that a classifier rates as risky. In a post on October 16, the team said quality held up. From June to September, merged pull requests more than doubled, and the share of them that had to be reverted fell by more than 40%.
The post on the Claude blog walks through the whole pipeline, and volume is what drove it. In Anthropic’s monorepo, changes landing per day rose from about 2,500 in May to more than 30,000 in September. Claude Code lives in its own repository. The team has said before that Claude writes about 90% of that code.
From a gate to spellcheck
Review used to be a gate at the end: write the code, open a PR, wait for a colleague. Now it runs while the code is being written, and most changes have been reviewed and fixed by Claude at least once before a human is asked to look. The target is for most pull requests in the Claude Code repos (CLI, desktop, web and mobile apps) to merge with at most one human in the loop, and that human can be the author. Reverts and incidents must not go up. The team tracks human approvals per PR at the median and the 95th percentile, along with human comments and how long a PR waits on a person. Revert and incident rates are the guardrail: if they rise, human time was only moved, not saved.
Much of the review now happens before a PR exists. Claude gets something it can run, such as tests, a build or a screenshot to compare against the design, and is asked to show the output. A subagent then reviews the diff with fresh context. Boris Cherny, who created Claude Code, estimates that verification makes the final result two to three times better: “Verification looks different for each domain. It might be as simple as running a bash command, or running a test suite, or testing the app in a browser or phone simulator.”
When the reviewer keeps flagging the same thing, the fix goes upstream: into a skill, a lint rule, CLAUDE.md, or a REVIEW.md that code owners write for their part of the code. The good REVIEW.md files are short and say what failure looks like in that repo. For docs it is “it lies”, for config “it misroutes”. They also cap nits and ask for a file and line as evidence.
How a finding gets posted
Every PR gets a Claude review that reads the diff against the whole repo and against what the PR says it is for. Several agents hunt for bugs in parallel. A separate group then votes on each finding, and a finding is posted only if enough of them agree it is real. Findings that are only about style, diagnostics or cleanup get dropped, a filter the team A/B tested on its own PRs. What remains arrives as inline comments tagged Important, Nit or Pre-existing, so a bug the PR didn’t introduce isn’t pinned on its author. The post ties a jump in the share of internal PRs getting substantive comments, from 16% to 54%, to the research preview launch. An Anthropic security post from July 26 gave the same figure and credited it to making the agents write a proof that each finding is valid.
A PR can’t merge while any finding from the bug review or the separate security review is still open. The author either fixes it or closes the thread with a reason, and if a later commit fixes the issue, the finding clears on its own. Then come lint, type checks, tests on every platform and a build. A typical CLI PR runs dozens of checks before anyone looks at it. The PR template asks how the change was tested and how to undo it. Before a change lands, the merge queue re-runs every check against main with the change applied. New behavior usually ships behind a flag to Anthropic employees first, then to the latest channel, with stable trailing behind.
Who has to approve?
The tier depends on the change, not on who wrote it. Low-risk changes (docs, copy, internal-only features, additive changes behind a flag) need no human approval once automated review and security checks are green. Ordinary product changes need one person, and so do sensitive areas such as caching, the auto-updater and system prompts. The highest-risk code and a small set of listed paths need two. One classifier picks the tier. A second asks whether this particular change needs another person even when its tier says it doesn’t. CI workflow files and anything owners have marked sensitive always need a human.
The lighter paths are only for authors on the repo’s trusted list, mostly the owning team. An agent’s approval never fills a human slot. On an agent-written PR, the person directing the agent counts as one of the required reviewers. An approval covers one specific commit, so an agent’s push or a hand-resolved merge means approving again. In July, the team described this split less formally: critical changes reviewed manually, with automated review increasingly covering the “outer layers”.
The fifty-round pull request
One CLI PR went through more than fifty rounds of agent review and fixes overnight before its author stopped it. The team ran git blame on what the reviewer kept flagging, and about 59% of it was code that had been there since the reviewer’s first pass. Each round found a few more old bugs. The team admits it had made this worse by running several rounds by default. Now there is a cap: after three rounds, the reviewer asks the author to step back and reconsider the approach. Later passes re-review only the new diff, findings name the class of bug so every instance gets fixed, and findings already posted are not repeated. Rounds-to-pass is tracked next to recall.
Noise matters more when the author is an agent. A person waves off a nit in seconds, but an agent fixes it, which means another commit and another round. The reviewer is scored daily against an eval built from past PRs that caused incidents, each bug confirmed by engineers who know the code. Each comment is also sorted by what the author did with it: fixed it, called it wrong, called the behavior intentional, or let it go. Noticing a bad merge still depends on the author, on-call and the revert and incident dashboards. There are no circuit breakers yet that tighten approval when those numbers move, and engineers want them before more code moves to zero-human merges.
People keep the calls machines are bad at: whether to build something at all, risky code such as billing, the auto-updater and CI workflows, architecture others will copy, reading the evidence, and owning the outcome. To learn a part of the codebase, engineers ask Claude for an explainer of that slice.
Our analysis
What we checked
On October 18 we ran the local /code-review on Claude Code 2.1.302, at default and high effort, against a scratch repo seeded with a real bug, a style nit, a pre-existing bug and a REVIEW.md. We also read the docs, CHANGELOG and the open-source security review Action.
In our runs the local command followed REVIEW.md well, but it tagged nothing by severity, and the first run said no verify pass ran, so the verification and tagging the post describes appear to belong to the managed review. The post also says the open-source security Action gets you the same thing as Anthropic’s gate. Going by that Action’s code, which we read but did not run, it only comments and never blocks a merge.
- Both runs reported 3 findings, opened with a count, gave file and line, and skipped the nits and
vendor/. Neither tagged anything Important, Nit or Pre-existing. The pre-existing February bug came back as an in-diff finding. - Per the CHANGELOG, at medium effort the local command also reports cleanup and CLAUDE.md-convention findings on models without tuned settings, including Opus 5.5 and Sonnet 5.5 (documented, not measured).
- The security Action posts its findings as a comment and never fails the check. By default it scans a PR only once, so nothing clears on its own (documented, not measured).
- The managed product never approves or blocks a PR, reads REVIEW.md only from the base branch, and bills each review as metered usage (documented, not measured).
Copying the setup
The post’s own adoption list starts small: put a short CLAUDE.md in place along with formatter and lint hooks, add a REVIEW.md, and turn on an AI reviewer for one repo for a few weeks before widening it. It also recommends writing down risk tiers that start strict, and using CODEOWNERS to make sure auth, billing, CI and installers always need a person. Since neither the managed review nor the Action blocks anything, the gate has to come from your own branch protection, for example by requiring all conversations to be resolved before merge.
Up next
Anthropic’s 45% quarter left IPO investors wanting more
Per Reuters, some Anthropic investors are weighing around $1.5 trillion for the IPO, down from roughly $2 trillion, after Q3 revenue grew 45%, slower than Q2’s doubling. In their place, would you buy in at $1.5 trillion or wait for a cheaper entry?
Related stories
- Claude Code team has Claude test the app before committing
- Claude’s merge rate outgrew git, so Anthropic built Scion
- Claude Code Review exits preview and runs about 3x faster
- Claude went from bug hunt to abuse report on New Year’s Eve
- Claude Code sessions keep running after you close the laptop
- Claude Code reads AGENTS.md when CLAUDE.md is missing
Comments
No comments yet. Be the first.
Join the conversation
Sign in to leave a comment. It appears after moderation.

