Products#Multi-agent#Coding agent
Anthropic published the Claude Code team's seven-layer review pipeline — and shipped Code Review to GA
Anthropic published its Claude Code team review pipeline on October 15 and shipped Code Review to GA, with at least $10,000 in credits for Enterprise.
On October 15, 2026, Anthropic published “How the Claude Code team does code review” on its claude.dev engineering blog, by Addy Osmani, Morgan Westlee Lunt and Qing Wang — the first full account of how the team that builds Claude Code reviews its own changes. The same day, Code Review, Anthropic’s pull-request reviewer for enterprises, left research preview and became generally available on Claude Team and Enterprise plans. The post puts numbers on the problem: Claude writes about 90 percent of Claude Code’s own code, and between June and September 2026 merged pull requests to the codebase more than doubled while the share that had to be reverted fell by more than 40 percent. Code Review is the productized version of that internal pipeline — an admin flips it on per repository and reviews run automatically on new PRs.
Key points
- The post numbers the pipeline 1 through 7. Human approval is tiered by risk: low-risk changes need no person once automated review and security checks are green, medium needs one, high needs two. The rules are “the change sets the tier, not the author” and “an agent’s approval never counts as a person’s.” An approval covers one specific commit; new agent commits need re-approval.
- Internal numbers: the share of internal PRs drawing substantive review comments went from 16% to 54% after Code Review was deployed. After the convergence work, PRs passing on one review-and-fix round rose from 65% to 78%, those needing five or more rounds fell from 8.9% to 1.5%, and mean reviews to pass fell from 2.08 to 1.32.
- The official announcement says Code Review is “out of research preview” for Team and Enterprise, that most reviews now finish about 3x faster than in preview (roughly 20 minutes down to 6-10), and that eligible Enterprise organizations can claim credits by January 15, 2027, expiring 30 days after they claim.
Background
Code Review launched as a research preview on March 9, 2026, for Claude for Teams and Claude for Enterprise customers. TechCrunch reported at the time that Anthropic’s head of product, Cat Wu, put the price at $15-25 per review billed on tokens, and described the architecture: multiple agents read the codebase in parallel from different angles, then a final agent aggregates, deduplicates and ranks the findings, focusing on logic errors rather than style. Anthropic also open-sourced a security-review GitHub Action during the preview; that repository now sits at about 6,400 stars.
The reason for the write-up is volume. Daily merges into Anthropic’s monorepo went from about 2,500 in May 2026 to more than 30,000 in September — numbers we covered in Anthropic ditched git for in-house Scion. The repository scopes are worth keeping straight: the post says Claude Code lives in its own repository, separate from the monorepo, so 30,000 is the monorepo’s figure, not Claude Code’s.
The same week offered a counterpoint. The kernel patches Linus Torvalds wrote with Claude Code were caught by a Meta maintainer in LKML review over a race condition and still have not been merged. Human review remains the last gate on the most expensive code, which is the same claim behind the post’s “the highest-risk code, plus a small set of paths that always need two people.”
The facts
In the post’s own numbering, the seven layers are: before the PR, Claude runs tests, build and linter, then a subagent reads the diff in fresh context; AI review on every PR; a separate security review that can block; automated checks that block (lint, type checks, tests on every platform, build); human approval scaled by risk; merge queue and staged release; and fast detection with fast rollback. Three mechanics are worth remembering. Several agents hunt for bugs at once, then a separate group votes on each finding and only agreed-upon ones get posted. Every finding must be fixed or closed with a reason. And findings are tagged Important, Nit or Pre-existing, so an old bug the PR did not introduce does not look like the author’s fault.
The human-approval tiers, from the post’s table:
| Risk | Examples (post’s wording) | Human approvals |
|---|---|---|
| Low | Docs, copy, internal-only features, additive changes behind a flag | None required once automated review and security checks are green |
| Medium | Ordinary product changes, and sensitive areas such as caching, the auto-updater or system prompts | One person |
| High | The highest-risk code, plus a small set of paths that always need two people | Two people |
The team also had to fix review that would not converge. The post says one CLI PR went through more than fifty rounds of review and fixes overnight before its author stopped it, and that when they ran git blame on what the reviewer kept flagging, about 59% was code that had been there since the reviewer’s first pass — it was finding a few more old bugs each round. They added three constraints: a round cap, after which the reviewer asks the author to reconsider the approach; diff-only re-review after the first few passes; and flagging the class of bug rather than one line. The table below is the internal before-and-after.
| Internal PR metric | Before | After |
|---|---|---|
| Pass after one review and fix round | 65% | 78% |
| Need five or more rounds | 8.9% | 1.5% |
| Mean reviews to pass | 2.08 | 1.32 |
On the product side, the docs are the authority. Admins set each repository to Quick, Standard (the default) or Deep, install the GitHub App and select repositories; reviews then run automatically on new PRs, and anyone with write access can trigger one with @claude review. Only an organization Owner can enable it, and drafts and fork PRs are skipped. The docs state three limits plainly: a review never approves or blocks the pull request, so the merge decision stays with the team; a review works from the diff and surrounding code without building or running the code; and it is not available self-hosted, or to organizations with zero data retention, the HIPAA configuration, or customer-managed encryption keys. Anthropic’s own repos go further than the product — the post says owners can let some low-risk changes merge on automated checks alone.
Cost is the item the docs refuse to pin down. The official costs page says Code Review bills on token usage through usage credits, separate from a plan’s included usage, and that it “has no flat per-review price”; spend varies with the review level, the size and complexity of the change, and the number of candidate issues needing verification. Organizations can set a monthly limit, and reviews skip once the budget is exhausted, leaving a comment on the PR explaining why. The GA announcement gives no current price, so the $15-25 figure from the preview should be read as history, not as today’s rate.
The credit offer is stated three different ways, so here is who says what. Anthropic’s announcement page says eligible Claude Enterprise organizations “can claim at least $10,000 in Code Review credits by January 15, 2027.” Anthropic’s support page says the amount is “between $10,000 and $50,000, based on your organization’s active Claude Code seats,” locked at the moment of claim. The Next Web’s headline says “$50,000 in credits attached” while its body gives “up to $50,000” and the $10,000-$50,000 range. Of the three, $50,000 is the top of a range rather than what every organization gets, and the announcement page states the floor. The support page also lists the exclusions: organizations billed through a reseller, those with a connected GitHub Enterprise Server or a data-residency GitHub Enterprise Cloud (.ghe.com) instance, trials and canceled subscriptions, Team/Pro/Max/Free plans, and anyone who claimed the earlier free offer.
What others say
The skepticism on Hacker News around this post concentrates on whether review still means anything at this scale; the thread’s score and comment count keep climbing, so we leave them out. User grey-area asks, “At 30k commits a day, is there anything that can meaningfully be called review?” mirekrusin calls it “conducting/steering/oversight/supervision, not reviewing.” markbao corrects a claim that spread in the thread: the 30,000 figure is the monorepo’s, and Claude Code lives in a separate repository — the user who first said “1,000 a day,” status_quo69, then conceded, “I did invent the 1000 a day because the article was vague.” User vavkamil argues that dismissing a critical finding as a false positive should involve AppSec, and points out that the README of Anthropic’s open-source security-review Action states the tool “is not hardened against prompt injection attacks and should only be used to review trusted PRs” — I read that sentence in the repository’s README.
The provenance of the post itself also came up. User paimapi (comment 50124937) says the detector Pangram reports 80% of the text is AI-generated and links the report, user typesanitizer (comment 50123523) writes verbatim that “Pangram identifies several sections as 100% AI-written,” and mk12 (comment 50124719) finds the prose style telling. Others push back — lavezzi suggests long immersion in a model’s output rubs off. Anthropic has not responded, so this stays a forum claim. Users ativzzz and sp4cec0wb0y both raise Goodhart’s law: once revert rates become the target, they stop being a good measure.
Independent reporting adds two things. The Next Web’s Ana Maria Constantin (2026-10-15) notes that European firms running GitHub data residency are excluded from the credits entirely, and recalls that in June 2026 a researcher showed Anthropic’s own Claude Code action carried a prompt injection flaw handing over an installation token with write access; Anthropic fixed the core bypass within four days, rated the issues 7.8 and paid a $4,800 bounty. She also reports that competitor CodeRabbit has raised $143M co-led by Atomico, runs two million reviews a week for 17,000 customers including BMW and Adyen, and that its CEO says OpenAI has replaced Anthropic as his main supplier. Unite.AI’s Jonas Reeve (2026-10-15) carries the credit range, the exclusions and a latent type-mismatch bug that early access caught in TrueNAS’s open-source middleware; the outlet’s byline discloses that the author is an AI research agent and the piece was reviewed by its editorial team.
One numeric disagreement is worth flagging: the Anthropic post says Claude writes about 90 percent of Claude Code’s own code, while The Next Web says 80 percent. Both describe the same claim, and we use the post’s 90.
Our take
The useful part of this post is not the announcement; it is the checklist you can copy. Give agents checks they can run (CLAUDE.md for the stack and conventions, hooks that run formatters and linters), require automated review and security review on every PR with every finding resolved before merge, write the risk tiers down and let the change rather than the author set them, never let a bot’s approval fill a human slot, catch surprises with a merge queue and feature flags, and turn past incidents into an eval set that runs daily. A team already letting agents open PRs can land all six this week.
The discount rate is high, and worth saying out loud. Every performance figure is Anthropic’s own; The Next Web puts it as “none of it has been independently tested,” and there is no external replication or control. The 16%-to-54% jump measures the share of PRs drawing substantive comments, not bug-detection rate, and that metric moves if reviewers simply post more comments. The docs themselves say reviews may miss bugs and security flaws and that a clean review does not mean the code is error-free. Neither the $10,000-$50,000 credit range nor the post-GA billing model gives a team a number it can budget against; the $15-25 per review from the preview is a historical data point. On security there is a plain admission to read first: Anthropic’s own open-source security-review Action says it is not hardened against prompt injection, which matters before wiring review into a workflow that accepts untrusted PRs. And the post admits a missing piece — there are no circuit breakers that tighten approval when revert rates move, which its own engineers want before more code moves to zero-human merges.
Who should care: teams already running Claude Code and letting agents open PRs, especially smaller ones stuck in a review queue. They can point @claude review or the local /code-review at a couple of PRs without changing any process, then decide whether to make review a blocking gate. Teams that want the headcount saving without maintaining a REVIEW.md and an eval set will mostly get a pile of comments that people wave off.
How to try it
Code Review requires a Team or Enterprise plan and repositories hosted on GitHub. An organization Owner enables it in Claude Code settings, installs the GitHub App and selects repositories; after that, reviews run automatically on new PRs with no developer configuration. To see the output without changing your process, anyone with write access can comment @claude review on a PR to trigger a single run. To run a review locally before you push, use the /code-review and /security-review commands in Claude Code, or wire up Anthropic’s open-source security-review GitHub Action (anthropics/claude-code-security-review) — read its prompt-injection note first.