Claude Code PRs doubled, reverts fell over 40%: Anthropic

The post frames the change as a shift in what the team measures. Rather than asking whether a human read the code, it asks what could break and which check, automated or human, is right to catch it. Human time becomes the cost to minimize, but only if the revert rate and incident rate stay flat, so the team treats those as guardrails against simply moving work elsewhere.

The review process itself is organized as overlapping layers rather than one perfect gate, with rollback as the floor beneath them. Human approval is one layer among several and is assigned by the risk of the change, not by who wrote it. A separate rule keeps agent approvals from counting as human ones, and assigns responsibility to whoever directed the agent.

The most practical passage concerns convergence between AI reviewers. One CLI PR ran through more than 50 review-and-fix rounds overnight before the author stopped it, and git blame showed about 59% of the flagged issues were already present when the reviewer first saw the code. The fixes, including a three-round cap and diff-only re-review, raised single-round merges from 65% to 78% and cut the average number of review rounds from 2.08 to 1.32.

What remains with people is a narrower set of judgments: whether to build something at all, high-risk code, architecture, reading evidence, and owning the result. The team says it keeps human review for changes from outside the team specifically so someone can ask whether the work should be done at all. It also says no one holds a full mental model of Claude Code, and that people learn a slice on demand instead.