Claude Code team doubled merged pull requests while reverts fell 40 percent

A team shipping code faster usually breaks more of it. The Claude Code team at Anthropic has published a long post saying it did the opposite over the summer, and names the four points in its flow where a person still has to read the diff.

Writing on claude.dev under the names Addy Osmani, Morgan Westlee Lunt and Qing Wang, the team reports that from June to September the number of pull requests merged into the Claude Code codebase more than doubled, and the share of merged pull requests that had to be reverted fell by more than 40 percent. The post also says Anthropic's main monorepo went from about 2,500 changes landing per day in May to more than 30,000 in September 2026, which the team attributes to agents writing most of the code.

The flow they describe

The post names six layers of checks on every change. Before a pull request exists, Claude runs the tests, build and linter, and a second agent reviews the diff in a fresh context. On every pull request, Anthropic's CI runs lint, types, tests across platforms, build and packaging, and red blocks the merge. A separate Claude code review runs in parallel, with agents hunting for bugs and verifiers voting on each finding. A security pass runs on top. A risk-based classifier then decides whether the change needs no human, one, or two.

After the merge, every change goes behind a flag, runs internally at Anthropic first, moves to the latest channel and then the stable channel, with the author and the on-call watching. Flags turn off quickly, broken builds get reverted.

The team's stated goal: most pull requests merge with at most one person in the loop, who can be the author, with no rise in reverts or incidents. The guardrail is reverts and incidents. If they rise, the time has moved, not fallen.

What they say Claude does differently

The authors say most review happens before the pull request. They cite Boris Cherny, who built Claude Code, as putting giving Claude a way to verify its work ahead of everything else, and quote him as estimating that it makes the final result two to three times better. The post also lists the places the team fixes a recurring reviewer finding: in a skill, a lint rule, or in CLAUDE.md, which is loaded on every review.

For a team shipping code with agents, the usable part of the post is the sequence rather than the headline. The number is Anthropic's own and on its own repo, so a team reading it should treat it as a case study. The sequence of checks and the single explicit guardrail of revert rate are what a team can copy today.

Source

How the Claude Code team does code review by Addy Osmani, Morgan Westlee Lunt and Qing Wang, Anthropic. Hacker News thread.