Claude Code Team's Approach to Code Review

How the team changed code review

Claude Code's team used to review code at the end of the process, when a person opened a pull request and waited for a colleague. Now review runs all the time, like spellcheck while you write. Claude checks most changes first, so a human only looks at the work after the routine mistakes are gone.

Key numbers

• From June to September, merged pull requests to Claude Code's codebase more than doubled.
• In the same period, the share of changes that had to be reverted fell by more than 40 percent.
• Changes landing per day in Anthropic's monorepo rose from about 2,500 in May to more than 30,000 in September.
• The team's goal: most pull requests need at most one human in the loop, with no rise in reverts or incidents.

Why the old way stopped working

When AI agents write code, the number of changes grows very fast. Reading every line by hand becomes almost impossible. So the team asks a different question: what could go wrong with this change, and which check, automated or human, is best for catching it?

Why it matters

Review has a cost in people's time, and so does fixing broken code. If the team only cuts review time, the time just moves to fixing bugs later. The team watches reverts and incidents to make sure human time is really saved. Their approach shows how a fast-moving team can use AI checks and still keep quality high.

What to watch

• Whether revert and incident rates stay flat as changes keep growing.
• How many human approvals and comments each pull request needs, and how long it waits for a person.
• How well Claude's self-checks, like running tests and comparing screenshots, catch problems before a human sees them.