# osmani-lunt-wang-code-review-claude-code-2026-10-15

## Veille

Engineering post by **Addy Osmani**, **Morgan Westlee Lunt**, and **Qing Wang** (the **Claude Code** team, **Anthropic**) on the *claude.dev* blog, published **October 15, 2026**; ~3,200 words. It describes the team's code review flow, in which Claude writes roughly **90%** of the code for Claude Code. From June to September, merged PRs **more than doubled** and the reverted share dropped by **more than 40%**; in Anthropic's monorepo, daily changes rose from **~2,500** (May) to **more than 30,000** (September).

## Titre Article

How the Claude Code team does code review

## Date

2026-10-15

## URL

https://claude.dev/blog/solving-code-review-at-anthropic/

## Keywords

code review, code review, agentic code review, Claude Code, Code Review for Claude Code, REVIEW.md, CLAUDE.md, risk-graded approval, risk classifier, security review, merge queue, feature flags, progressive rollout, rollback, incident eval, reviewer noise, review convergence, round cap, human time, revert rate, author accountability, code explainer, verification

## Authors

Addy Osmani, Morgan Westlee Lunt, Qing Wang — équipe Claude Code (Anthropic) ; blog *claude.dev*, catégorie Engineering.

## Ton

Team debrief, in the first-person plural, in a factual and operational register. The text poses a question (how to review at the volume where agents produce), redefines the purpose of review, then follows the life of a change layer by layer, before describing the reviewer's improvement, the place of humans, and the spread of context. It draws on internal figures (volumes, revert rate, share of commented PRs, review rounds) and five diagrams. It flags its limits: no *circuit breakers* yet that tighten approval when indicators shift, no formal process for sharing context, one PR that went through more than fifty rounds. It closes with a staged adoption list and a call for feedback from teams that let agents open PRs. The Anthropic products cited (Code Review, `/code-review`, `/security-review`, GitHub Action) are presented as the building blocks available.

## Pense-betes

- **Goal**: the question is no longer *"did a human read the code?"* but *"what can go wrong, and what check should catch it?"*. Review becomes continuous, like a spell checker.
- **Metrics**: human approvals per PR (median, p95), human comments, wait time on a person; safeguard = revert rate and incident rate.
- **Before the PR**: Claude shows the output (tests, build, screenshot); a subagent reviews the diff in a fresh context; internal skills for bug hunting, simplification, E2E. Boris Cherny estimates that verification makes the result **2 to 3 times** better.
- **REVIEW.md**: written by the code owners; short; states what failure looks like (docs: *"it lies"*, config: *"it misroutes"*), caps nits, requires file and line as evidence.
- **AI review**: PRs with substantive comments **16% → 54%**; Quick / Standard / Deep levels; verifier voting; Important / Nit / Pre-existing tags; every finding must be addressed.
- **Tiers**: low (docs, flag) = zero humans; medium = one; high = two. The change sets the tier, not the author; an agent's approval never replaces a human's; approval tied to a commit.
- **Convergence**: 59% of repeated flags were about code present since the first pass. After fixes: one-round pass **65% → 78%**, five rounds or more **8.9% → 1.5%**, average **2.08 → 1.32**.
- **"You land it, you own it"**: small changes, everything reversible, the impact surface sets the requirement, no blame.
- **Shared context**: a Claude-generated explainer of the code slice to be modified rather than reading trivial diffs.
- Neighbor of [[houck-dx-what-are-code-reviews-for-2026-08-05]] and [[sfeir-code-review-anneau-contraintes-2026-07-30]].

## RésuméDe400mots

The Anthropic Claude Code team describes how it reviews code that is roughly 90% produced by Claude. Volume has exploded: in Anthropic's monorepo, daily changes rose from about 2,500 in May to more than 30,000 in September, and manual line-by-line review is no longer sustainable. From June to September, merged PRs on Claude Code more than doubled while the reverted share dropped by more than 40%.

Review serves two functions: catching defects and spreading knowledge of the code. For the first, the team replaces the question "did a human read it?" with "what check should catch what can go wrong?". The goal: most PRs merge with at most one human, with no increase in reverts or incidents, which act as a safeguard.

A change passes through several layers. Before the PR, Claude runs tests, build, and linter and shows the output, then a subagent reviews the diff in a fresh context. CLAUDE.md describes the project; REVIEW.md, written by the code owners, tells the reviewer what matters. On each PR, several agents hunt for bugs in parallel, others vote on each finding, and only the retained findings are published, sorted by severity. A separate security review and CI block the merge as long as an issue remains open. A classifier sets the risk level, and therefore the number of human approvers (zero, one, or two); some paths always require a person, and an agent's approval never counts as a human's. Merge queue, feature flags, internal use first, a latest channel ahead of stable, and fast rollback round out the setup.

The reviewer improves in a loop. PRs that caused incidents, confirmed by experts, form an eval replayed every day. Noise is measured by how authors react, since an agent fixes every nit and reruns a round. One PR went through more than fifty rounds in a single night; the team capped rounds at three, limited re-review to the new diff, and removed repeats, bringing the average down from 2.08 to 1.32 rounds.

Humans keep five decisions: whether to build it, risky code, architecture, reading the evidence, and accountability for the outcome. To understand the code, Claude is asked for an explainer of the relevant slice. The "you land it, you own it" norm relies on changes that are small and reversible. The article closes with a staged adoption list.

## GrapheDeConnaissance

- Addy Osmani —publie→ article How the Claude Code team does code review (DOCUMENT, 0.98)
- Morgan Westlee Lunt —publie→ article How the Claude Code team does code review (DOCUMENT, 0.98)
- Qing Wang —publie→ article How the Claude Code team does code review (DOCUMENT, 0.98)
- Anthropic —mesure→ changements quotidiens du monorepo : ~2 500 en mai → plus de 30 000 en septembre 2026 (MESURE, 0.95)
- Claude Code —mesure→ PR fusionnées ×2+ et part revertée −40 %+ de juin à septembre 2026 (MESURE, 0.95)
- Anthropic —affirme_que→ « un humain a-t-il lu le code ? » est un proxy imprécis de « quel contrôle doit attraper ce qui peut mal tourner ? » (AFFIRMATION, 0.93)
- revue de code agentique —réduit→ temps de revue humain (CONCEPT, 0.9)
- revue de code agentique —utilise→ Code Review for Claude Code (TECHNOLOGIE, 0.92)
- revue de code agentique —utilise→ REVIEW.md (DOCUMENT, 0.92)
- revue de code agentique —utilise→ CLAUDE.md (TECHNOLOGIE, 0.9)
- revue de code agentique —utilise→ approbation graduée par le risque (METHODOLOGIE, 0.92)
- revue de code agentique —utilise→ /security-review (TECHNOLOGIE, 0.88)
- Code Review for Claude Code —mesure→ PR internes avec commentaires substantiels : 16 % → 54 % (MESURE, 0.93)
- approbation graduée par le risque —utilise→ classifieur de risque (TECHNOLOGIE, 0.9)
- eval d'incidents —améliore→ Code Review for Claude Code (TECHNOLOGIE, 0.88)
- plafond de rondes de revue —réduit→ itérations revue-correction (CONCEPT, 0.9)
- plafond de rondes de revue —mesure→ rondes moyennes avant passage : 2,08 → 1,32 ; cinq rondes ou plus : 8,9 % → 1,5 % (MESURE, 0.93)
- Boris Cherny —affirme_que→ donner à Claude un moyen de vérifier son travail rend le résultat 2 à 3 fois meilleur (AFFIRMATION, 0.9)
- Anthropic —recommande→ ne jamais laisser l'approbation d'un bot remplacer celle d'un humain requis (AFFIRMATION, 0.93)
- Anthropic —recommande→ mesurer le temps humain avec taux de revert et d'incident comme garde-fou (AFFIRMATION, 0.92)
- You land it, you own it —s_applique_à→ Claude Code (TECHNOLOGIE, 0.88)
- Claude Code —utilise→ merge queue (TECHNOLOGIE, 0.88)
- Claude Code —utilise→ feature flags (TECHNOLOGIE, 0.88)
- article How the Claude Code team does code review —référence→ Boris Cherny (PERSONNE, 0.92)

---
Canonical: https://www.thekb.eu/en/fiches/osmani-lunt-wang-code-review-claude-code-2026-10-15/