Industry#Coding agent#Code review
Anthropic published the Claude Code team review process: doubled merges, 40% fewer reverts
Anthropic's engineering post details the Claude Code team's review pipeline: merged PRs doubled June–September, reverts fell 40%, reviews cap at 3 rounds.
Anthropic published its own team’s code-review pipeline on October 15, 2026, in the engineering post “How the Claude Code team does code review”: Claude writes about 90% of Claude Code’s own code, merged PRs more than doubled from June to September, and the share of merged PRs that had to be reverted fell by more than 40%. The pipeline stacks six layers of checks — from pre-submit verification to risk-tiered human approval — with fast rollback underneath. After Addy Osmani shared the post on X, it passed a million views within a day and became the reference text in this week’s engineering debate over AI code review.
Key points
- Claude writes about 90% of Claude Code; from June to September 2026 merged PRs “more than doubled” and the reverted share “fell by more than 40 percent” (official blog figures).
- The pipeline runs six layers: pre-PR verification, required CI, AI code review, security review, risk-tiered human approval, and merge queue plus feature flags for gradual release — with fast rollback as the recovery path.
- Reviews cap at three rounds; before the cap, one CLI PR went through more than 50 rounds overnight.
Background
As AI-written code volume climbs, review has replaced generation as the bottleneck: an agent can submit a dozen PRs overnight, and human line-by-line review neither scales nor knows where to look. Anthropic’s answer is to automate the reviewing too while keeping the merge decision human. The post lives on the product site’s Engineering section (claude.dev), not the main company blog, credited to Addy Osmani, Morgan Westlee Lunt and Qing Wang. Independent verification appeared quickly: the author of the paddo.dev blog ran a “second model” review script that produced 597 review reports between September 20 and October 27, re-checking changes Claude had already approved.
The facts
Per the Anthropic engineering post (October 15, 2026):
| Item | Official figure |
|---|---|
| Claude-written share of Claude Code | ~90% |
| Merged PRs, June–September | more than doubled |
| Reverted share of merged PRs | fell by more than 40% |
| Review-round cap | 3 rounds, then rethink the approach |
| Human approvals by risk | low 0 / ordinary 1 / highest 2 |
| Outlier case | one CLI PR through 50+ review rounds (pre-reform) |
Risk tiers key off the change, not the author: docs and copy — low risk — need no human approval once automated checks are green; ordinary product changes (including sensitive areas like caching and the auto-updater) take one approver; the highest-risk code and designated paths take two. Supporting rules: only trust-listed authors get the lighter paths, an agent’s approval never counts as a human’s, approvals bind to a specific commit and must be redone on unverifiable new commits, and the person prompting an agent-written PR counts as a required reviewer. Security review is its own layer, where an open finding blocks the merge; the AI-review layer runs parallel bug-hunter subagents over the diff, with agent voting to filter false positives.
What others say
paddo.dev’s “Claude Reviewing Claude” is the sharpest external check: its second-model script re-reviewed 51 changes Claude had already passed, found confirmed bugs in 45, and 29 of those were serious — while about a quarter of the second model’s findings were false positives. The critique there targets the homogeneity rather than the pipeline: every reviewer in Anthropic’s stack is Claude itself, and “if Claude has a blind spot, five Claudes share it,” with Codex, Grok and Gemini catching serious bugs Claude passed (double-refunded Stripe payments, an unauthenticated local control server, a path-traversal bug) as the running counterexample.
Osmani’s X share pushed the post past a million views in a day, and discussion concentrated on whether zero-human-approval for low-risk changes is aggressive. Practitioner threads on HN mostly converged on the three-round cap: broad agreement that review rounds don’t correlate with quality, but worry that a hard cap pushes edge cases into side channels.
Our take
For teams using AI to write code, three pieces here transfer directly: cap review rounds (the “more review is better” instinct is measurable and wrong), tier by the change rather than the author (a senior name waives nothing), and demand evidence before approval (an agent saying “done” counts for nothing until tests, builds and screenshot comparisons pass). Read the honest limits too: the 40% revert reduction is Anthropic’s own number with no third-party baseline, the six-layer pipeline has been validated inside a hundred-person-scale engineering org and would likely be over-engineering for a team of five, and the structural risk of an all-Claude review stack is already quantified by paddo.dev’s 29 serious bugs — if your production code is largely generated by a single model, putting a heterogeneous model in the review line is the move the post doesn’t make but its own data supports. For the failure mode of trusting a single model’s self-report, see our earlier piece on Cammarata’s “already clean” interpretability claim and the second opinions it drew.