• Sources: primary
  • Summary: Reading Anthropic's post, Theo reports that changes landing per day in Anthropic's monorepo rose from 2,500 in May to more than 30,000 in September, and that merged Claude Code PRs more than doubled from June to September while the revert share fell by more than 40 percent. Each PR passes CI, a multi-agent bug review with voting, a blocking security review, and a risk classifier that requires zero, one or two humans, and an agent approval never counts as a human one. Capping review rounds and re-reviewing only diffs cut mean rounds to pass from 2.08 to 1.32 and PRs needing five or more rounds from 9 to 1.5 percent, and Theo adds his own case of an uncapped agent review loop that ran four days and grew a PR to about 80,000 lines.
  • Why it matters: It is a mechanism-level template for AI pre-review, risk-based human gates and review-loop caps that a team can copy, with the metrics Anthropic uses as guardrails.
  • Background: In agent-driven workflows a coding agent writes the change and other agents review it, so review can loop indefinitely when the author agent fixes every nit a reviewer agent raises. Anthropic's answer is to measure human time per PR and to treat revert and incident rates as guardrails, so saved review time that shows up as more incidents does not count as a gain.