Anthropic: Claude writes 90% of Claude Code, reviewed in 7 layers

Anthropic says Claude generates about 90% of Claude Code's own code, with merged PRs more than doubling from June to September while reverts fell over 40%. Review runs through seven layers: pre-PR agent checks, parallel AI bug hunting with voting, a separate security review and risk-tiered human approval.

• Merged PRs more than doubled from June to September, reverts fell over 40%
• PRs passing after one review cycle rose from 65% to 78%
• Share of PRs needing 5+ review rounds dropped from 8.9% to 1.5%
• After three rounds the reviewer must ask the author to rethink the approach