Anthropic says Claude now writes about 90% of Claude Code, with low-risk changes merged after automated checks

Anthropic said about 90% of the code in Claude Code is now written by Claude, and that it has reworked code review to match the speed of AI-generated output.

From June to September this year, the Claude Code team more than doubled the number of merged pull requests, while the share of changes that later had to be reverted fell by more than 40%.

A layered review process built around automated checks

According to Anthropic, Claude runs tests before each submission and then has an independent sub-agent inspect the change. After the code is submitted, multiple agents search for bugs in parallel, and another group of agents cross-checks those findings. Security review and automated testing can also stop problematic code from being merged.

Human approval still depends on risk level

Human review has not been removed. Instead, Anthropic said approvals are assigned by risk category. Low-risk changes, including documentation edits from trusted developers, can be merged directly after passing automated checks. Standard product changes require one human approval, while high-risk code requires two.

The company also said AI approval cannot replace human approval, and developers remain responsible for the final code.

Anthropic says AI review can fall into loops

Anthropic disclosed that one Claude Code change went through more than 50 rounds of review and fixes overnight. After examining the case, the team found that about 59% of the issues involved code that was already present in the first review round, rather than problems introduced by later edits.

It then changed the rules. If a change goes through more than three review rounds, the author is asked to reconsider the approach, and repeated feedback is limited to keep agents from getting stuck with one another.