Anthropic published how the Claude Code team reviews code, and capped its own reviewer at three rounds

On October 15 Anthropic published how the Claude Code team does code review. Agents write the code on Jon's own site, and a reviewer reads each change before it ships. Anthropic's post is about that same job run at volume, and it includes the night its reviewer would not stop and the rules that fixed it. The headline figure is the company's own and comes with no counts: from June to September, merged pull requests more than doubled while the share that had to be reverted fell by more than 40 percent.

The failure Anthropic describes is one pull request that "went through more than fifty rounds of review and fixes overnight before its author stopped it." About 59 percent of what the reviewer kept flagging was code that had been there since its first pass.

The same day, the Code Review product became generally available for Team and Enterprise plans. Most reviews now finish in 6 to 10 minutes, and an admin sets each repository to Quick, Standard or Deep. The product reads and comments. Per its documentation, it works "without building or running your code" and "never approves or blocks the pull request." It has "no flat per-review price" and is billed on tokens. The Next Web reports that reviews ran $15 to $25 each during the preview, a price only that outlet reports, and that none of Anthropic's figures have been independently tested.

Two things make this more than a product launch. The reviewer is one of seven layers on that team, and the 40 percent came from all seven: checks that run code, a security review, human approval scaled by risk, a merge queue and fast rollback. And an agent that writes and an agent that reviews can keep each other busy with no end, which on token billing is a bill as well as a delay.

How to use it

Most of it needs no purchase:

1. Run before review. Give the agent tests, a build or a screenshot to compare, and have it show the output.
2. Cap the rounds. After three, the author steps back and rethinks the approach. After the first few passes, re-review only the new diff. Flag the class of bug so every instance is fixed at once. Never repeat a finding already posted.
3. Set risk from the change. Their tiers: docs, copy and additive changes behind a flag need no person once the checks are green. Ordinary product changes need one. The riskiest paths need two. "An agent's approval never counts as a person's."
4. Write a REVIEW.md. Say what failure looks like in that repository, cap the nits, list paths to skip, and ask for a file and line as evidence.
5. Add "Testing", "Risk" and "How to revert" to the pull request template.
6. Turn reverts into the reviewer's test. Anthropic reruns its reviewer daily against the pull requests that caused incidents.

In Claude Code, /code-review and /security-review run the same kind of pass locally.

The product itself is for Team and Enterprise plans: an admin turns it on in Claude Code settings, installs the GitHub App and picks repositories. The catch: a repository starts on "After every push", which multiplies the cost by the number of pushes, and a review with no findings "doesn't mean your code is error-free or secure."

What nobody knows yet

• The counts behind the 40 percent, and what else changed in those four months.
• Whether the process holds for a small team without a lab's token budget, the main objection on Hacker News.
• Anthropic names its own gap: it has no circuit breakers that tighten approval when reverts climb.

What to watch

• An outside team publishing its own revert numbers with Code Review turned on.
• January 15, 2027: the last day eligible Enterprise organizations can claim Code Review credits.