Claude Code begins having AI review AI: 90% of code written by Claude, low-risk changes require no human approval

Beating AI News Flash: Anthropic disclosed that about 90% of Claude Code's code is now written by Claude. From June to September this year, the number of merged code change requests (PRs) by the Claude Code team more than doubled, while the proportion requiring reversal actually dropped by more than 40%.

To keep up with the speed of AI writing code, Anthropic restructured its code review process. Before each submission, Claude first runs tests and has an independent sub-agent review the changes. After submission, multiple agents search for bugs in parallel, and then another group of agents cross-validates. Security reviews and automated tests also block problematic code from being merged.

Human review is tiered by risk. For low-risk changes such as documentation edits submitted by trusted developers, they can be merged directly after passing automated checks; ordinary product changes require approval by 1 person, and high-risk code requires 2 people. AI approval cannot replace human review, and developers remain responsible for the final code.

However, AI review can also get stuck in an infinite loop. Anthropic disclosed that one Claude Code change once went through more than 50 rounds of review and fixes overnight. A team inspection found that about 59% of the issues involved code that already existed during the first round of review, rather than new problems introduced by later modifications. The team then adjusted the rules: after more than 3 rounds of review, the author is required to reconsider the approach, and repeated feedback is limited to prevent agents from getting entangled with each other.