The AI Record
Review has to converge. An agent author and an agent reviewer can keep each other busy indefinitely. One CLI PR went through more than fifty rounds of review and fixes overnight before its author stopped it. When we ran git blame on what the reviewer kept flagging, about 59% of it was code that had been there since the reviewer's first pass.

AnthropicPublished by the organization itself (post), no named author

Post: How the Claude Code team does code review, claude.dev Blog (Anthropic), October 15, 2026, at How we make the reviewer better.

  • 50 rounds review rounds single pr, one night, one Claude Code CLI pull request. No person in the loop
  • 59% repeat findings share

Checked against the source [ Grade B ] Said in the last year Anthropic: $47 billion revenue, run-rate, May 2026 Sells AI tools

Filed under

Anthropic sells AI or agent tools or services, so it has a reason to publish flattering numbers about AI. The passage stays in the archive; read it with that in mind.

How this was checked

Checked against the source by a first reader, then confirmed independently by a second reader who fetched the source without seeing the first reader's notes (Independent review agent (Claude Sonnet), blind second pass from passage text and citation only, no repository access, 2026-11-07).

Under 'How we make the reviewer better': 'Review has to converge. An agent author and an agent reviewer can keep each other busy indefinitely. One CLI PR went through more than fifty rounds of review and fixes overnight before its author stopped it.' The git blame sentence ('about 59% of it was code that had been there since the reviewer's first pass') also matches. The 'Review has to converge.' lead-in is bold in the page markup, but the text is the same.

How to cite this passage

Anthropic, “How the Claude Code team does code review”, claude.dev Blog (Anthropic), 2026-10-15, at How we make the reviewer better. https://claude.dev/blog/solving-code-review-at-anthropic/ Archived at https://the-ai-record.com/passages/anthropic-one-pr-fifty-review-rounds-overnight-2026

More from the same source

The change sets the tier, not the author. A classifier reads the change and picks its risk tier. … An agent's approval never counts as a person's, either. Automated review can clear a low-risk change, but it can't fill a slot that calls for a human.

AnthropicPublished by the organization itself (post), no named author

Context. From the section on human approval, where the post sets three risk tiers for changes to Claude Code's repos, needing no, one or two human approvals.

Post: How the Claude Code team does code review, claude.dev Blog (Anthropic), October 15, 2026, at Layer by layer: 5. Human approval, scaled by risk.

Checked against the source [ Grade B ] Said in the last year Anthropic: $47 billion revenue, run-rate, May 2026 Sells AI tools Passage page

From June to September, merged pull requests to Claude Code's codebase more than doubled while the share that had to be reverted fell by more than 40 percent. We didn't remove review. But we did move most of the reading to machines, optimized for human time and attention, and invested heavily in making it fast and easy to undo mistakes.

AnthropicPublished by the organization itself (post), no named author

Post: How the Claude Code team does code review, claude.dev Blog (Anthropic), October 15, 2026, at Introduction.

  • 2x output change, June to September 2026. Level not stated
  • 40% revert rate reduction, June to September 2026. Level not stated

Checked against the source [ Grade B ] Said in the last year Anthropic: $47 billion revenue, run-rate, May 2026 Sells AI tools Passage page

So the goal is that most pull requests in the Claude Code repos … merge with at most one human in the loop, who can be the author, with no increase in reverts or incidents. To measure human time, we track human approvals per PR (median and 95th percentile), human comments, and how long a PR waits on a person. Revert and incident rates are the guardrail: if they rise, we haven't saved human time, we've only moved it.

AnthropicPublished by the organization itself (post), no named author

Post: How the Claude Code team does code review, claude.dev Blog (Anthropic), October 15, 2026, at The scarce resource: human time and attention.

Checked against the source [ Grade B ] Said in the last year Anthropic: $47 billion revenue, run-rate, May 2026 Sells AI tools Passage page

We think "did a human read the code?" is an imprecise proxy for "what could go wrong with this change, and which check—automated or human—is the right one to catch it?"

AnthropicPublished by the organization itself (post), no named author

Post: How the Claude Code team does code review, claude.dev Blog (Anthropic), October 15, 2026, at Rethinking the purpose of code review.

Checked against the source [ Grade B ] Said in the last year Anthropic: $47 billion revenue, run-rate, May 2026 Sells AI tools Passage page

We measure noise by watching what people do with the comments. It's easy to catch more bugs if you don't mind flagging a lot of non-bugs, so we need a check on that. We sort every comment the reviewer leaves by how the person on the PR responded: they fixed it, said it was wrong, said the behavior was intentional, or decided it didn't matter. We also run the reviewer on PRs we know are clean and count how often it reports something that would have led someone to make a harmful change.

AnthropicPublished by the organization itself (post), no named author

Post: How the Claude Code team does code review, claude.dev Blog (Anthropic), October 15, 2026, at How we make the reviewer better.

Checked against the source [ Grade B ] Said in the last year Anthropic: $47 billion revenue, run-rate, May 2026 Sells AI tools Passage page

We don't yet have circuit breakers that tighten approval when those numbers move; engineers here want them before more code moves to zero-human merges.

AnthropicPublished by the organization itself (post), no named author

Context. "Those numbers" are the revert and incident rates: the sentence before says noticing a bad merge still depends on the author, on-call, and the revert and incident dashboards.

Post: How the Claude Code team does code review, claude.dev Blog (Anthropic), October 15, 2026, at Layer by layer: 7. Fast detection, fast rollback.

Checked against the source [ Grade B ] Said in the last year Anthropic: $47 billion revenue, run-rate, May 2026 Sells AI tools Passage page

Any incidents are now an eval. The team collected past PRs that caused incidents, with their fix-forwards and reverts, and had engineers who know that code confirm each bug and where it lived. The reviewer runs against that set daily.

AnthropicPublished by the organization itself (post), no named author

Post: How the Claude Code team does code review, claude.dev Blog (Anthropic), October 15, 2026, at How we make the reviewer better.

Checked against the source [ Grade B ] Said in the last year Anthropic: $47 billion revenue, run-rate, May 2026 Sells AI tools Passage page

Pulling a senior engineer out of deep work to catch a routine mistake costs far more than an automated review. People on the team are still responsible for: … Reviewing risky code such as billing, the auto-updater and CI workflows. … Owning the outcome. A person is accountable for every change that ships, whoever typed it.

AnthropicPublished by the organization itself (post), no named author

Context. The two cuts drop three items from the post's list of what people stay responsible for: deciding whether to build a change, reviewing architecture that sets a pattern others will copy, and reading the evidence the checks produce.

Post: How the Claude Code team does code review, claude.dev Blog (Anthropic), October 15, 2026, at Where humans stay in the loop.

Checked against the source [ Grade B ] Said in the last year Anthropic: $47 billion revenue, run-rate, May 2026 Sells AI tools Passage page

Before and after the convergence work: pull requests passing after one review and fix round rose from 65% to 78%; pull requests needing five or more rounds fell from 8.9% to 1.5%; mean reviews to pass fell from 2.08 to 1.32.

AnthropicPublished by the organization itself (post), no named author

Context. The text description of a chart in the post comparing internal pull requests before and after changes to re-review, which the post lists as a cap of three rounds, re-reviewing only the new diff after the first few passes, flagging the class of bug, and not repeating findings already posted.

Post: How the Claude Code team does code review, claude.dev Blog (Anthropic), October 15, 2026, at How we make the reviewer better: Figure D, Review that converges (the chart's text description).

  • 78% prs pass after one review round, internal pull requests, after the convergence changes (0.65 before). Level not stated
  • 1.5% prs needing five plus review rounds, internal pull requests, after the convergence changes (0.089 before). Level not stated
  • 1.32 iterations per PR avg review iterations, internal pull requests, after the convergence changes (2.08 before). Level not stated

Checked against the source [ Grade B ] Said in the last year Anthropic: $47 billion revenue, run-rate, May 2026 Sells AI tools Passage page

Noise costs more when the author is an agent. A person can wave off a nit in seconds. An agent fixes it, which means another commit and another round of review.

AnthropicPublished by the organization itself (post), no named author

Post: How the Claude Code team does code review, claude.dev Blog (Anthropic), October 15, 2026, at How we make the reviewer better.

Checked against the source [ Grade B ] Said in the last year Anthropic: $47 billion revenue, run-rate, May 2026 Sells AI tools Passage page

Findings get verified by other agents before humans review them. Several agents look for bugs at the same time. Then a separate group votes on each finding, and it only gets posted if enough of them agree it's real. Findings that are only about style, diagnostics, or cleanup get dropped. We A/B tested those filters on our own PRs before keeping them.

AnthropicPublished by the organization itself (post), no named author

Post: How the Claude Code team does code review, claude.dev Blog (Anthropic), October 15, 2026, at Layer by layer: 2. AI review on every pull request.

Checked against the source [ Grade B ] Said in the last year Anthropic: $47 billion revenue, run-rate, May 2026 Sells AI tools Passage page

When we launched the research preview, the share of internal PRs getting substantive review comments went from 16% to 54%. Most PRs over 1,000 lines get findings, and most small ones come back clean.

AnthropicPublished by the organization itself (post), no named author

Context. About Code Review for Claude Code, Anthropic's AI pull-request reviewer, which the post says is now generally available on its Team and Enterprise plans.

Post: How the Claude Code team does code review, claude.dev Blog (Anthropic), October 15, 2026, at Layer by layer: 2. AI review on every pull request.

  • 54% prs with substantive review comments, internal pull requests, after the Code Review research preview launched (0.16 before). AI assists a person

Checked against the source [ Grade B ] Said in the last year Anthropic: $47 billion revenue, run-rate, May 2026 Sells AI tools Passage page

Something wrong? Report a correction. JSON: /api/v1/passages/anthropic-one-pr-fifty-review-rounds-overnight-2026.json