How the Claude Code team does code review / claude.dev Blog
The Claude Code team explains how it reviews code when agents write about 90 percent of it, moving most of the reading to automated reviewers and spending human attention where machines can't catch the problem. The core idea is to minimize human review time without raising reverts or incidents.
- Claude reviews its own diffs before a PR exists, and every PR gets an automated review that verifies findings with a separate voting group before posting them, dropping style-only noise.
- Review is tiered by risk: low-risk changes need no human approval once automated checks pass, medium-risk changes need one person, and the highest-risk paths need two, with a classifier setting the tier rather than the author.
- From June to September, merged PRs more than doubled while the share needing revert fell by over 40 percent, which they present as evidence the approach works.
The Claude Code team's core idea is that "did a human read the code?" is the wrong question. The better one is what could go wrong with a given change, and which check (automated or human) is best placed to catch it. Their goal is to minimize human review time per pull request without raising reverts or incidents. Review now runs continuously instead of as a gate at the end: Claude reviews its own work before a PR exists, and most routine mistakes get caught before a person opens the diff. Between June and September, merged PRs more than doubled while the share needing revert fell by more than 40 percent. Volume is the pressure behind all this. Changes landing daily in Anthropic's monorepo jumped from about 2,500 in May to over 30,000 in September, and Claude writes roughly 90 percent of Claude Code's own code.
The mechanics are concrete. Authors give Claude something it can run, like tests, a build, or a screenshot to compare against a design, and make it show output rather than claim success. A fresh-context subagent reviews the diff. Recurring review findings get fixed at the source, through skills, lint rules, or project files: CLAUDE.md describes the stack and conventions, and per-repo REVIEW.md files written by code owners say what failure looks like in that codebase. Every PR gets an automated review against the whole repo and the stated intent. Multiple agents hunt for bugs, a separate group votes on each finding, and only findings that clear that bar get posted, tagged Important, Nit, or Pre-existing. When the team launched the preview, the share of internal PRs receiving substantive comments rose from 16 percent to 54 percent. Security findings block merge until resolved or dismissed with a reason.
Human approval scales with risk. Low-risk changes like docs, copy, and additive flagged features need no human approval once automated checks pass. Ordinary and sensitive changes (caching, the auto-updater, system prompts) need one person. The highest-risk code, plus a small set of paths, needs two. The change itself sets the tier, not the author's seniority. The team tracks human approvals per PR, comments, and wait time as its efficiency measures, with revert and incident rates as the guardrail. If the guardrail moves up, they've only shifted the cost from review into cleanup.