How the Claude Code team does code review

Anthropic's Claude Code team published a detailed engineering post on October 15, 2026, explaining how it reviews code at the scale that AI agents produce it — and the answer involves removing humans from most of the path. According to claude.dev's report, the team now routes the vast majority of pull-request review through layered automated checks, with human attention reserved for decisions only people can make.

This piece covers the full review system Anthropic built for Claude Code's own codebase — the six-layer pipeline, the AI-driven review tooling now available to paying customers, the risk classifier, and the rollback infrastructure that makes the whole thing safe to run at speed. It is aimed at engineering leads and platform teams thinking about how to keep code quality high as AI-generated commits multiply.

What shipped in Claude Code team does code review

Anthropic's Claude Code team shipped a production-grade, six-layer code review pipeline for its own CLI, desktop, web, and mobile repositories, and made the core review capability — Code Review for Claude Code — generally available on Team and Enterprise plans. The public version reads each pull request diff against the full repository and the PR's stated intent, checking whether the change solves its stated problem, fits the broader system, and avoids breaking anything else. A lighter local variant runs via the /code-review slash command inside Claude Code.

The system processes changes before a PR exists, during CI on every PR, at a security-review layer, through a risk-based approval classifier, inside a merge queue, and then again after merge behind feature flags. Most of the path is automated. Anthropic employees run new builds internally before they reach outside users, followed by a staged release — the latest channel first, the stable channel trailing it — with flags that can be turned off quickly and broken builds reverted by the author and on-call team.

What improved in Claude Code team does code review

Between June and September 2026, merged pull requests to Claude Code's codebase more than doubled while the share that required reversion fell by more than 40 percent. Across Anthropic's monorepo, changes landing per day climbed from roughly 2,500 in May to more than 30,000 in September. The team's goal — that most PRs merge with at most one human in the loop, who can be the author — was achieved without raising revert or incident rates.

When Anthropic launched the Code Review research preview, the share of internal PRs receiving substantive review comments jumped from 16 percent to 54 percent. Most PRs over 1,000 lines generate findings; most small ones come back clean. Boris Cherny, who created Claude Code, estimates that giving Claude a way to verify its own work makes the final result two to three times better.

What you gain from Claude Code team does code review

Teams adopting the same pipeline get three concrete benefits: fewer human review hours spent on routine diffs, earlier discovery of bugs before a PR even exists, and a verified finding mechanism that filters AI hallucinations before they reach engineers. Multiple agents hunt for bugs in parallel on every PR; a separate group of verifier agents then votes on each finding, and a result is only posted to the PR if enough voters agree it is real. Style and nit comments that pass the relevance bar but are not bugs are kept separate and capped per REVIEW.md configuration.

The risk-based approval classifier assigns each PR to one of three tiers: no human required, one approver, or two approvers. Certain code paths always require a person regardless of tier. Admins can configure each repository to Quick, Standard, or Deep review modes; Anthropic's guidance is that Deep is worth the extra time when one missed bug costs more than the review itself. That tiering lets teams concentrate human attention on high-risk changes while letting low-risk, clean PRs flow through with no waiting.

How to get Claude Code team does code review

Code Review for Claude Code is generally available today on Team and Enterprise plans. To update Claude Code to the latest version and access the feature:

claude update

To run a local code review on the current diff inside any Claude Code session:

/code-review

To launch Claude Code pinned to the model used by the Claude Code team's review agents:

claude --model claude-code-review

Repository admins configure the review depth per repo — Quick, Standard, or Deep — through the Claude Code admin settings for their organization. CLAUDE.md files at the repo root set project-wide conventions that every review loads automatically; REVIEW.md files, written by code owners, tell the reviewer what matters in their specific part of the codebase, cap the nit count, list paths to skip, and require file-and-line citations rather than guesses from symbol names.

What to watch after Claude Code team does code review

The Claude Code team flags two open problems. The first is codebase understanding: line-by-line review traditionally builds shared context across a team, and moving most reading to machines risks eroding that context over time. The team acknowledges this explicitly and says it will return to the question, but has not yet described a replacement mechanism. Engineering leads rolling out similar pipelines should track whether junior engineers lose the code-reading exposure that review once provided.

The second is the verification loop at the pre-PR stage. Cherny's estimate — that verifiable output makes results two to three times better — assumes teams invest time writing runnable checks, skills, and good CLAUDE.md and REVIEW.md files. Teams that skip that investment and rely only on the automated PR-level review will see a narrower improvement. The revert-rate guardrail the Claude Code team uses — measuring revert and incident rates alongside human approvals per PR — is a useful adoption signal for any team trying to calibrate how far to push automation.