Claude Code team: merged PRs more than doubled while reverts fell by more than 40 percent

The authors say Claude writes about 90 percent of Claude Code's own code. From June to September, merged pull requests to its codebase "more than doubled while the share that had to be reverted fell by more than 40 percent." The review path has layers: the author's agent verifies its own work and a second agent reviews the diff in fresh context; every PR gets a Claude review in which several agents hunt for bugs and a separate group votes on each finding; a security review can block the merge; and human approval is set by a risk classifier, with none, one or two people required. Open findings from either review block the merge, and an agent's approval never counts as a person's.

The post also reports a failure of its own. One CLI PR went through more than fifty rounds of review and fixes overnight. By the authors' count, about 59% of what the reviewer kept flagging was code present since its first pass. After adding a three-round cap, diff-only re-review and no repeated findings, they report that PRs passing after one round rose from 65% to 78%, PRs needing five or more rounds fell from 8.9% to 1.5%, and mean reviews to pass fell from 2.08 to 1.32. They also say past incidents and reverts now form a daily eval set for the reviewer, and that they do not yet have circuit breakers that tighten approval when revert or incident rates rise. All figures are the team's own internal measurements; the post publishes no underlying data.

Why it matters: The useful part is the shape of the loop: risk-tiered human approval, a reviewer scored against your own incidents, and a cap on review rounds so an agent author and an agent reviewer do not keep each other busy. The numbers come from one vendor's own repository, so treat them as a design reference, not a benchmark.