How Anthropic’s Claude Code Team Reviews Code

Anthropic’s Claude Code engineering team has published a detailed account of how it reviews code when agents write most of it. In “How the Claude Code team does code review” on claude.dev (15 October 2026), Addy Osmani, Morgan Westlee Lunt and Qing Wang say merged pull requests to Claude Code’s codebase more than doubled from June to September while the share that had to be reverted fell by more than 40 percent. Those are Anthropic’s own figures, not an independent audit.

They also restate an earlier claim: “We’ve said before that Claude writes about 90 percent of Claude Code’s own code.”

The scarce resource is human attention

The team’s stated goal is that most pull requests in the Claude Code repos (CLI, desktop and web apps, mobile apps) “merge with at most one human in the loop, who can be the author, with no increase in reverts or incidents.” Review, they write, now “runs continuously, more like spellcheck while you write,” with most changes reviewed and fixed by Claude at least once before a human looks.

They did not remove review. “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.”

Background: volume across Anthropic

As background (already covered in TSN’s Scion / monorepo story), the same post notes that in Anthropic’s monorepo, changes landing per day went from about 2,500 in May to more than 30,000 in September. Claude Code lives in its own repository, separate from that monorepo, but faces similar pressure. The new news for this post is the Claude Code merge/revert trend and the review stack, not the monorepo volume.

Layers of review

The post describes a multi-layer path rather than a single end gate:

Pre-PR verification — give Claude runnable checks (tests, build, screenshot vs design); subagent review with fresh context; internal skills for bugs, simplification and end-to-end behaviour; CLAUDE.md and owner-written REVIEW.md.
AI review on every PR — Claude reads the diff against the repo and stated intent; findings verified by voting agents before humans see them; tags Important / Nit / Pre-existing. The public product, Code Review for Claude Code, is generally available on Team and Enterprise; /code-review is a lighter local version.
Security review that can block — open findings block merge until fixed or dismissed with a reason.
Automated CI — lint, types, tests, builds; PR template asks how it was tested and how to undo it.
Human approval scaled by risk — low (docs, flagged additive): none once automated/security green; medium: one person; high / sensitive paths: two people. An agent’s approval never counts as a person’s. Trusted authors only get lighter paths.
Merge queue, flags, staged release — re-check against main; internal-first flags; latest then stable.
Fast rollback — flags off without deploy; on-call; incident reviews ask which layer should have caught the issue.

When the research preview of Code Review launched, the post says the share of internal PRs getting substantive review comments went from 16% to 54%.

Making the reviewer converge

Agent-author plus agent-reviewer loops can thrash. The team describes a PR that went through more than fifty rounds overnight before the author stopped it, and git blame work suggesting about 59% of repeat flags were code present since the first pass. Changes include a three-round cap, diff-only re-review after early passes, flagging a class of bug so all instances get fixed, and no repeats of findings already posted. Incidents and reverts become a daily eval set for the reviewer.

Where humans stay

Humans keep: whether to build it at all; risky areas (billing, auto-updater, CI); architecture that sets patterns; reading evidence (tests, screenshots, findings); and owning the outcome. “You land it, you own it” is the stated norm for small and medium changes, backed by small diffs, flags and fast reverts.

Why it matters

This is a primary-source playbook for reviewing agent-written code at a frontier lab that also sells the reviewer. The June–September merge-up / revert-down claim is the measurable claim to watch — and to treat as company-reported until outsiders can see the same dashboards. For teams adopting Claude Code Review, the useful exports are risk tiers, round caps, REVIEW.md, and never letting a bot fill a required human approval slot.

What’s still unknown

Absolute PR and revert counts (only relative claims published)
Whether revert/incident rates stayed flat outside Claude Code repos
How far “zero-human merge” for low-risk paths will extend

What this means

Anthropic (company figures): Claude Code merged PRs more than doubled June–September; share reverted fell more than 40%.
Restated: Claude writes about 90% of Claude Code’s own code.
Process: continuous automated review; humans for risk, taste and ownership; goal ≤ one human per PR without more reverts/incidents.
Monorepo volume (2,500→30,000+/day) is background only — already on TSN via Scion.