Anthropic Reveals 90% of Claude’s Code Was Written by AI, Streamlining the Review Process

Dynamic Watch Beating AI News: Anthropic revealed that approximately 90% of Claude Code’s code has been written by Claude. Between June and September this year, the number of code change requests (PRs) merged by the Claude Code team more than doubled, while the percentage requiring rollback decreased by over 40%. To keep pace with AI-generated code, Anthropic restructured its code review process. Before each submission, Claude runs tests and assigns independent sub-Agents to review the changes. After submission, multiple Agents concurrently search for bugs, followed by cross-verification by another set of Agents. Security reviews and automated testing also prevent problematic code from being merged. Human reviews are tiered by risk level: low-risk changes, such as documentation updates from trusted developers, can be merged directly after automated checks; standard product changes require approval from one person; high-risk code requires approval from two. AI approvals cannot replace human judgment—developers remain responsible for the final code. However, AI reviews can sometimes enter loops. Anthropic disclosed that one Claude Code change underwent over 50 rounds of review and repair in a single night. Upon investigation, the team found that about 59% of issues stemmed from code that already existed during the first review, not from newly introduced changes. The team subsequently adjusted its rules: after more than three review rounds, authors must reconsider their approach, and repetitive feedback is restricted to prevent Agents from getting stuck in circular loops.