CodingSoftware & developer tools
Anthropic's Claude Code team lets low-risk changes merge without human approval and describes its seven-layer review process
With Claude writing about 90% of Claude Code and daily monorepo changes rising from 2,500 to over 30,000, Anthropic's team now uses a classifier to sort changes into risk tiers. Low-risk changes merge without human approval once automated and security checks pass.
From the Simon Willison article of October 16, 2026 and cited sources: Human approval now depends on risk tier: a change needs zero, one or two approvers. [2] [1]
SOURCES[2] How the Claude Code team does code review (claude.dev) claude.dev[1] Simon Willison: Quoting Addy Osmani, Morgan Westlee Lunt… simonwillison.net
AT Pulse research deskPublished · Updated · 5 min read · AI-assisted analysisAI-assisted, not human-edited. Methodology Corrections Who owns this site
- What happened
- Human approval now depends on risk tier: a change needs zero, one or two approvers. [2] [1]
- Original source
- Simon Willison article, independent perspective
- Brief by
- AT Pulse research desk AI-assisted; publisher and contact on the desk page
- When
- October 16, 2026 article published
- Covered
- Oct 17 edition
- Evidence
- 8 sources across 8 sites
- 1Simon Willison: Quoting Addy Osmani, Morgan Westlee Lunt…simonwillison.net
- 2How the Claude Code team does code review (claude.dev)claude.dev
- 3heise: Anthropic relieves developers: Code review with AI…heise.de
- 4Bringing Code Review to Claude Code (Anthropic)claude.com
- 5InfoQ: Claude Code Reviewinfoq.com
- 6Unite.AI: Anthropic Makes Code Review for Claude Code…unite.ai
Reach: not measured yet
Human approval now depends on risk tier: a change needs zero, one or two approvers. That step is one of seven layers, which also include a self-check with a fresh-context subagent, AI and security review, required CI, a merge queue that re-checks changes, and fast rollback. [2] [1]
The authors argue that asking whether a human read the code is a weak substitute for asking what could go wrong. Review becomes a continuous process, similar to spellcheck, rather than a single gate. [1] [2]
Anthropic reports that from June to September, merged PRs more than doubled while the share of reverts fell by more than 40%. These are internal figures with no published baseline method. PRs that passed review in one round after fixes rose from 65% to 78%. [2]
The managed Code Review product is now generally available. Reviews take about 6–10 minutes. Parallel agents look for issues, a verification step filters false positives, and findings are ranked by severity. The reviewer has read-only access and never approves or blocks a PR. It is limited to Team and Enterprise plans on GitHub and is not available with Zero Data Retention, HIPAA configuration or customer-managed keys. [7] [6]
The evidence is one team's internal process and metrics.
Read the full assessment
The implication is that as AI-generated code outgrows what people can read, scrutiny may move toward risk classification, required CI, merge queues and fast rollback rather than human review of every line.
Executive brief
Anthropic says Claude writes about 90% of Claude Code's own codebase. Changes per day in its monorepo went from 2,500 in May to more than 30,000 in September 2026. The team reports that merged PRs more than doubled from June to September while the share of reverts fell by more than 40% (claude.dev).
Read the full section
Anthropic says Claude writes about 90% of Claude Code's own codebase. Changes per day in its monorepo went from 2,500 in May to more than 30,000 in September 2026. To keep up, the Claude Code team stopped requiring a human to read every change. A classifier now sorts each change into a risk tier, and low-risk changes merge with no human approval once automated and security checks pass. The team reports that merged PRs more than doubled from June to September while the share of reverts fell by more than 40% (claude.dev). The figures are internal, and no one outside Anthropic has checked them. Simon Willison highlighted the post's main idea: review that runs continuously, "like spellcheck."
What changed and event timeline
Code Review enters research preview
Anthropic launched multi-agent PR review for Team and Enterprise. Reviews took about 20 minutes and cost an estimated $15–25 each (;).
Developers raise cost and circularity concerns
InfoQ reported doubts about whether the per-review price was worth it, and objections to Claude reviewing code that Claude wrote ().
Monorepo volume jumps about 12x
Daily changes in Anthropic's monorepo rose from 2,500 to more than 30,000, according to the team ().
Process write-up and general availability
Osmani, Lunt and Wang published the team's seven-layer review process. Code Review became generally available, with reviews about 3x faster (6–10 minutes) (;).
Willison amplifies the core argument
His quotation post singles out the authors' point that asking whether a human read the code is an imprecise stand-in for asking what could go wrong ().
Capabilities and access
- Code Review (managed): Team and Enterprise plans, GitHub only (including GitHub Enterprise Server).
- Models: Anthropic picks them and admins cannot choose. The launch used Claude 4-series models (heise). No exact version is published.
- Review levels: Quick, Standard or Deep, set per organization or per repository (docs).
Read the full section
- Code Review (managed): Team and Enterprise plans, GitHub only (including GitHub Enterprise Server). It is not available with Zero Data Retention, HIPAA configuration or customer-managed encryption keys (docs).
- Models: Anthropic picks them and admins cannot choose. The launch used Claude 4-series models (heise). No exact version is published.
- Review levels: Quick, Standard or Deep, set per organization or per repository (docs).
- Other options: local
/code-reviewon any plan, Ultrareview, or GitHub Actions or GitLab CI with your own API key. - Launch credits: eligible Enterprise organizations can claim $10k–$50k in credits through Jan 15, 2027 (Unite.AI).
Technical analysis for researchers and developers
- Pipeline: parallel agents look for issues, a verification step drops false positives, then findings are deduplicated and ranked by severity. Access is read-only.
- Prompt-injection guard:
REVIEW.mdis read only from the base branch, never from the PR's own changes. - Seven internal layers: pre-PR self-check plus a fresh-context subagent, AI review, security review, required CI, risk-tiered human approval (0, 1 or 2 approvers), a merge queue with re-verification, and fast rollback (claude.dev).
Read the full section
- Pipeline: parallel agents look for issues, a verification step drops false positives, then findings are deduplicated and ranked by severity. Access is read-only. The reviewer reads the diff and surrounding code but never builds or runs it, and it never approves or blocks a PR (docs).
- Prompt-injection guard:
REVIEW.mdis read only from the base branch, never from the PR's own changes. - Seven internal layers: pre-PR self-check plus a fresh-context subagent, AI review, security review, required CI, risk-tiered human approval (0, 1 or 2 approvers), a merge queue with re-verification, and fast rollback (claude.dev).
- Feedback loop: incidents feed into daily evaluation sets for the reviewer.
- Convergence: after fixes, PRs passing in one review round rose from 65% to 78%. Average rounds to pass fell from 2.08 to 1.32.
Claims and evidence
- Claude writes about 90% of Claude Code, and monorepo changes reached 30k+ per day (claude.dev).
- Merged PRs more than doubled and the revert share fell more than 40% (Jun–Sep). These are Anthropic-only numbers with no published baseline method.
- Substantive review comments rose from 16% to 54% of PRs. On PRs over 1,000 lines, 84% got findings, averaging 7.5 issues each (InfoQ).
Read the full section
- Claude writes about 90% of Claude Code, and monorepo changes reached 30k+ per day (claude.dev).
- Merged PRs more than doubled and the revert share fell more than 40% (Jun–Sep). These are Anthropic-only numbers with no published baseline method. No independent corroboration was found, and third-party write-ups repeat the claim without testing it (CellCog, which also discloses that it builds on Claude).
- Substantive review comments rose from 16% to 54% of PRs. On PRs over 1,000 lines, 84% got findings, averaging 7.5 issues each (InfoQ).
- Engineers marked fewer than 1% of findings as incorrect (Unite.AI). That measures what engineers flagged, not a measured false-positive or missed-bug rate.
Context and prior work
- AI-assisted review competes with GitHub Copilot code review and CodeRabbit. Anthropic pitches deeper, slower analysis (InfoQ).
- heise framed human review as the new bottleneck once AI makes code output much larger (heise).
- Anthropic presents automated review as one layer of its secure development lifecycle, not a replacement for it (docs).
Limitations, safety and contested findings
- Anthropic's own docs warn that a review with no findings does not mean the code is error-free or secure (docs).
- The team admits it lacks "circuit breakers" that would automatically tighten approval rules when revert or incident rates move.
- Others question cost against open-source alternatives (InfoQ).
Read the full section
- Anthropic's own docs warn that a review with no findings does not mean the code is error-free or secure (docs).
- The team admits it lacks "circuit breakers" that would automatically tighten approval rules when revert or incident rates move. The model depends on feature flags, staged releases and on-call rollback (claude.dev).
- Critics call same-vendor write-and-review circular. Others question cost against open-source alternatives (InfoQ).
- The reviewer does not execute code, so runtime-only bugs depend on CI.
Business and practitioner implications
- Copy the risk tiers first, not the tool. Dropping human approval only works alongside required CI, merge queues, feature flags and fast rollback.
- Preview reviews cost $15–25 each, Deep reviews cost more, and spend limits exist (Unite.AI).
- Track human approvals per PR, wait times on people, reverts and incidents, as the Claude Code team does.
Read the full section
- Copy the risk tiers first, not the tool. Dropping human approval only works alongside required CI, merge queues, feature flags and fast rollback.
- Budget for it. Preview reviews cost $15–25 each, Deep reviews cost more, and spend limits exist (Unite.AI).
- Measure the guardrails. Track human approvals per PR, wait times on people, reverts and incidents, as the Claude Code team does.
- Check compliance before adopting. Regulated organizations using Zero Data Retention, HIPAA or customer-managed keys are currently excluded.