Skip to content- Sources: primary, discussion
- Summary: From June to September, merged PRs to Claude Code more than doubled and the share that had to be reverted fell by more than 40 percent. Every PR gets a Claude review whose findings other agents vote on before posting, plus a blocking security review. A classifier assigns each change a risk tier needing zero, one or two human approvals, and only trusted authors get the lighter paths. After review loops reached more than fifty overnight rounds, the team capped reviews at three rounds and switched to diff-only re-review.
- Why it matters: It gives a concrete template for risk-tiered approval, reviewer evals built from past incidents, and round caps that keep agent author and reviewer loops from running forever.