GitHubWhat they did
Addy Osmani, Morgan Westlee Lunt and Qing Wang describe how the Claude Code team reviews code now that Claude writes about 90 percent of it. Before a PR exists, Claude is asked to run tests, a build or a screenshot comparison and show the output, then a subagent with fresh context reviews the diff, and authors run internal skills that hunt bugs, simplify the diff and check behaviour end to end. Every PR then gets a Claude review in which several agents look for bugs and a separate group votes on each finding before it is posted inline, tagged Important, Nit or Pre-existing. A security review can block the merge. A classifier sorts each change into a low, medium or high risk tier that sets whether zero, one or two people must approve, and agents often pick up review findings overnight. Review agents also sweep existing code nightly.
How it went
From June to September merged PRs to the Claude Code codebase more than doubled and the share that had to be reverted fell by more than 40 percent. When the reviewer launched internally, PRs with substantive review comments went from 16 to 54 percent. Most PRs over 1,000 lines get findings and most small ones come back clean. One CLI PR went through more than fifty overnight rounds of agent review and fixes before its author stopped it, and about 59 percent of repeat flags were old code, so they capped re-review at three rounds and switched to diff-only passes.
Worth knowing
This is Anthropic writing about its own product, so the numbers are self-reported. Past incident PRs became a daily eval for the reviewer, and noise is measured by how authors respond to each comment. HN commenters were sceptical of change counts as a metric, and one pointed out that the open-source security review action says it is not hardened against prompt injection.