Claude Code hands most code reading to machines: merged pull requests doubled while reverts fell more than 40 percent

Code review shifted from humans reading every line to machines screening first, with human time reserved for judgments only people can make. Review used to be a gate at the end of the road, after the code was written and a colleague read the pull request line by line; now it runs continuously like spellcheck, and most changes have been reviewed and fixed by Claude at least once before a human is asked to look. From June to September, merged pull requests to the Claude Code codebase more than doubled while the share that had to be reverted fell by more than 40 percent.

Every pull request gets a Claude review before a person looks, reading the diff against the whole repository and the request's stated intent to ask whether the change solves its problem, fits the system, or breaks something else. When the research preview launched, the share of internal pull requests getting substantive review comments went from 16 percent to 54 percent; most pull requests over 1,000 lines get findings, and most small ones come back clean. Several agents look for bugs at the same time, then a separate group votes on each finding and it is posted only if enough agree it is real; findings only about style, diagnostics, or cleanup are dropped, and those filters were A/B tested on the team's own pull requests.

The number of human approvals is set by the change's risk tier rather than by its author. Low-risk changes such as docs and copy need no human approval once automated review and security checks are green; ordinary product changes and sensitive areas such as caching, the auto-updater, or system prompts need one person; the highest-risk code and a small set of paths always need two. A classifier reads the change and picks its risk tier, and a second classifier asks whether this particular change needs another person even when its tier says it does not; only authors on the repository's trusted list get the lighter paths, and an agent's approval never counts as a person's.

The process covers Claude Code's own repositories, the CLI, desktop and web apps, and mobile apps, and its authors are an internal Anthropic team. It measures human approvals per pull request, human comments, and how long a pull request waits on a person, with revert and incident rates as the guardrail. The adoption steps target teams willing to let agents open pull requests, starting on one repository and tuning REVIEW.md for a few weeks before widening.

The team does not yet have circuit breakers that tighten approval when revert or incident numbers move, and engineers want them before more code moves to zero-human merges. Review convergence remains open: one CLI pull request went through more than fifty rounds of review and fixes overnight before its author stopped it, and git blame on what the reviewer kept flagging showed about 59 percent was code that had been there since the reviewer's first pass. The team also says it has no formal process for spreading codebase understanding once machines do most of the reading.