How Anthropic reviews code that Claude writes 90% of itself

Anthropic published a detailed account of how its own Claude Code team handles code review, and the headline number is that Claude now writes roughly 90% of Claude Code’s own codebase. According to the post itself, authored by Addy Osmani, Morgan Westlee Lunt, and Qing Wang, that volume shift is dramatic by any measure: daily changes landing in Anthropic’s monorepo went from around 2,500 in May to more than 30,000 in September, a scale of churn that would bury a traditional human review process within days. Between June and September specifically, merged pull requests more than doubled while the share that had to be reverted fell by more than 40%, the two numbers the team treats as proof the system is actually working rather than just moving faster.

The review pipeline itself runs seven layers deep before anything ships, starting before a pull request even exists. Claude runs its own tests, builds, and linters, and a separate subagent reviews the diff with fresh context, catching issues before a human would ever see them. Every PR then gets a parallel AI review, where multiple bug-hunting agents flag potential problems and separate verifier agents vote on whether each finding is real, tagging results as Important, Nit, or Pre-existing so nothing gets silently dismissed. A standalone security review runs alongside that process, and open findings from either pass block the merge outright. Required CI, a merge queue that reruns checks against the latest main branch, feature flags for gradual rollout, and a monitored rollback path round out the rest of the pipeline.

Humans haven’t been removed from the loop, they’ve been triaged into it based on risk rather than inserted into every change by default. A classifier reads the change itself, not who wrote it, and sorts it into one of three tiers: low-risk changes like documentation or additive features behind a flag can merge with zero human sign-off once automated checks are green, standard product changes need one person’s approval, and the highest-risk code, plus a fixed list of sensitive paths like CI workflow files, always needs two. An agent’s approval never counts toward that requirement, and when a human directs an agent to write a PR, that person is automatically on the hook as a required reviewer rather than being able to wave the whole thing through untouched.

Anthropic has turned the internal version of this into an actual product rather than keeping it in-house: Code Review for Claude Code is now generally available on Team and Enterprise plans, configurable per repository as Quick, Standard, or Deep, alongside a lighter /code-review command anyone can run locally and a separate /security-review flow with an open-source GitHub Action. During the research preview leading up to this, the team says the share of internal pull requests getting a substantive review comment, something beyond a rubber stamp, rose from 16% to 54%. For teams running Claude Code at any real scale, that’s the actual pitch here: not that AI writes most of the code, but that a tiered, agent-checked review process is what makes shipping that much AI-written code survivable without burying every PR in human review time.