Anthropic has published how the Claude Code team reviews changes now that Claude writes roughly 90% of its code. A classifier assigns a risk tier, low-risk changes need no human approval, medium-risk ones need one person, and an agent's approval never counts as a human's. One pull request went through more than 50 rounds of review and fixes overnight, so the team capped the number of rounds. InfoQ also quotes Hacker News commenters, one who said nobody holds a causal model to debug from, and another whose SOC 2 reviewer required a second person to see every change.
