Anthropic's Claude Code Team Scales Code Review With AI-First Process

Anthropic's Claude Code team more than doubled merged pull requests from June to September while the share reverted fell by more than 40%. Every PR now gets a Claude review first, and the number of required human approvals scales with each change's risk. The reviewer reads each diff against the whole repo and the PR's stated intent.

Anthropic's Claude Code team more than doubled the number of pull requests merged to Claude Code's codebase from June to September, while the share that had to be reverted fell by more than 40 percent. Every pull request now gets a review from Claude before a person looks at it.

The team laid out its process in a post on claude.dev published Thursday, Oct. 15. Anthropic has said Claude writes about 90 percent of Claude Code's own code, and the post takes up how to review code at that volume. "We didn't remove review," the authors wrote. "But we did move most of the reading to machines, optimized for human time and attention, and invested heavily in making it fast and easy to undo mistakes."

Code review does two jobs, according to the post: it finds and fixes issues before they ship, and it builds shared context and understanding of the codebase across the team. The post mostly covers the first. The team calls "did a human read the code?" an imprecise proxy for a better question: what could go wrong with this change, and which check, automated or human, is the right one to catch it?

Claude reads each pull request's diff against the whole repository and the PR's stated intent, then asks whether the change solves its problem, fits the system or breaks something else. On a bug fix, it looks for the root cause. The public version, Code Review for Claude Code, is now generally available on Team and Enterprise plans, and the /code-review command runs a lighter version locally. When the research preview launched, the share of internal PRs getting substantive review comments rose from 16% to 54%.

In the public product, several agents examine the diff and surrounding code in parallel, a verification step filters out false positives, and the remaining findings are deduplicated, ranked by severity and posted as inline comments with a summary and a check run, according to Anthropic's documentation. That product never approves or blocks a pull request. Anthropic's own repos are stricter: a PR can't merge while any finding from the bug review or a separate automated security review is still open, and each finding has to be fixed or resolved with a reason.

Human approval is scaled by risk. Low-risk changes such as docs, copy and additive changes behind a flag need no human approval once automated review and security checks are green. Ordinary product changes need one person, and the highest-risk code needs two. A classifier sets the tier from the change rather than the author, and an agent's approval never counts as a person's. People now concentrate on direction and scope, risky code such as billing and the auto-updater, and architecture that others will copy. The goal is the minimum amount of human review time that doesn't raise reverts or incidents.

The pressure extends well beyond Claude Code. In an engineering post published Tuesday, Oct. 13, Anthropic said changes merged to its monorepo went from about 2,500 a day in May to more than 30,000 a day in September. To keep up, it built Scion, an in-house version control system, along with a code review tool called Antique. That review tooling can express semantic constraints on approvals that separate comment-only changes from sensitive logical changes, so human reviewers can focus on the most consequential ones. The source control team still reviews every change request for Scion, but its members now examine the intent of a change instead of reading every line.

The Claude Code team has also had to stop agents from reviewing in circles. One CLI pull request went through more than fifty rounds of review and fixes overnight before its author stopped it. Now, after three rounds, the reviewer asks the author to step back and reconsider the approach.