Anthropic's Claude Code team runs a six-layer code review process in production across the Claude Code repositories (the CLI, the desktop and web apps, and the mobile apps). Per the team's 15 October 2026 engineering blog post, the goal is that most pull requests merge with at most one human in the loop, who can be the author, with no increase in reverts or incidents; human time is the scarce resource, tracked via human approvals per PR (median and 95th percentile), human comments, and PR wait time on a person, with revert and incident rates as the guardrail.
The six layers are: (1) pre-PR verification by Claude itself with a runnable check (tests, build, or a screenshot to compare against design), plus a fresh-context subagent review and internal skills for bugs, diff simplification, and end-to-end behavior; (2) AI review on every PR against the whole repo and the PR's stated intent, with multiple agents independently looking for bugs and a separate group voting on each finding before any inline comment posts; (3) a security review that blocks merge while any finding is open; (4) the usual automated checks (lint, type, tests, build); (5) human approval scaled by risk tier, with the tier set by a classifier reading the change rather than the author (Low = no human required once automated review is green, Medium = one person, High = two people, plus a small set of paths that always need a person); (6) a merge queue that re-runs every check against main with the change applied, with new behavior shipping behind flags to Anthropic employees first and to public `latest` before `stable`. A seventh layer covers fast detection and rollback: flags turn off without a deploy, and a weekly on-call rotation runs incident reviews asking which layer should have caught the issue.
Quantified outcomes reported in the post, all on the Claude Code codebase between June and September 2026: merged pull requests more than doubled while the share that had to be reverted fell by more than 40 percent; the share of internal PRs getting substantive review comments went from 16 percent to 54 percent after launch of the research preview of Code Review for Claude Code. In the wider Anthropic monorepo, changes landing per day went from about 2,500 in May to more than 30,000 in September 2026. Boris Cherny, who created Claude Code, estimates that giving Claude a way to verify its work (running a bash command, a test suite, or a browser or phone simulator) makes the final result two to three times better.
The post calls out three specific operational practices that govern how the process is run: a short CLAUDE.md loaded by every review and pointed at from the project root; a per-repo REVIEW.md written by code owners that asks the reviewer for a file and line as evidence instead of a guess from a name; and an "approval covers one specific commit" rule, where any new commit that the system cannot vouch for (an agent's push, a manually-resolved merge) requires the human approval to be given again. Agent approvals never count as a person's, and on agent-written PRs the person directing the agent counts as one of the required reviewers the same as if they had typed the code themselves.