How Anthropic reviews code when Claude Code writes 90% of it

A new kind of code review

When AI agents write the majority of a project's code, traditional pull-request review breaks down. At Anthropic, the team behind the Claude Code AI coding assistant found itself in exactly that situation. Claude now writes about 90 percent of Claude Code's own code, and the volume of changes has been staggering.

Between May and September, daily changes landing in Anthropic's monorepo rose from roughly 2,500 to more than 30,000. Claude Code lives in its own repository but feels the same pressure. The team's answer was not to slow down, but to rethink what code review actually does and shift the work accordingly.

Moving review earlier, not skipping it

Code review traditionally serves two purposes: catching bugs before they ship, and helping the team share understanding of the codebase. The Claude Code team focused on the first, and changed when review happens.

Rather than waiting for a pull request and then asking a colleague to read every line, most review now runs continuously, more like spellcheck while you write. By the time a human looks at a change, Claude Code has usually reviewed and fixed it at least once. This catches routine mistakes early and saves team attention for harder questions.

"Verification looks different for each domain," said Boris Cherny, who created Claude Code. "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." He estimates that giving the agent a way to verify its own work makes the final result two to three times better.

Layers instead of gates

Most changes now travel through several automated layers before reaching a person. Before opening a pull request, many engineers use internal review skills that hunt for bugs, simplify diffs, and check behavior end to end. A subagent then reviews the diff with fresh context. UI changes are also checked against design guidelines.

The team keeps project expectations explicit in files such as CLAUDE.md, which summarizes the stack, conventions, goals and non-goals, and REVIEW.md, written by code owners to explain what matters in their part of the code. Good REVIEW.md files are short, describe what failure looks like, cap nitpicks, and ask for concrete evidence rather than guesses.

Every pull request still receives a Claude Code review before any human sees it. The agent reads the diff against the full repository and the PR's stated intent.

Measuring success by reverts, not approval speed

The goal is simple in theory and hard in practice: most pull requests should merge with at most one human in the loop, without increasing reverts or incidents. To track this, the team monitors how long pull requests wait on a person, how many human comments appear, and how often changes have to be reverted.

Between June and September, the number of combined pull requests merged into Claude Code's repository more than doubled, and the percentage that needed to be reversed dropped by over 40 percent. The team stresses that it did not remove review. It shifted the bulk of reading work to machines, prioritizing human time and focus, and made significant investments in ensuring mistakes could be undone quickly and easily.

Key facts

• Claude writes about 90 percent of Claude Code's own code
• Daily changes at Anthropic climbed from roughly 2,500 in May to over 30,000 in September
• Merged pull requests doubled while revert rates fell by more than 40 percent between June and September
• Giving Claude a way to verify its own work makes results two to three times better, according to Boris Cherny

Sources

1. How the Claude Code team does code review / claude.dev Blog

Every fact in this story comes from the source above.