In-depth analysis · About 5 min read
Claude Code's review process reduces human reading while keeping human responsibility
Anthropic describes layered checks, risk-based approval and faster rollback, with limits on what its public review product actually does.
- Source published
- Published here
- Source checked
In an October 15 engineering account, the Claude Code team explains how it reviews agent-written changes without requiring a colleague to read every line. The approach moves routine checking earlier, varies human approval by risk and measures the cost of mistakes as well as time spent reviewing. It does not remove accountability or turn an AI review into a human approval.
Listen to the news
Read this article aloud or keep listening to the news.
Preparing the reader…
Device voice
Speech is handled by your browser or device. AISubHub does not send the text to an audio API. Voice availability and playback with a locked screen vary by device.
Resuming may repeat a short phrase.
At a glance
- From June to September, merged pull requests in the Claude Code codebase more than doubled while the share needing reversion fell by more than 40%, according to the team.
- Trusted low-risk changes may need no human approval after checks pass; ordinary changes need one and the highest-risk changes two. Required humans cannot be replaced by agents.
- The public Code Review product, now generally available on Team and Enterprise plans, posts findings rather than approvals; the team's internal workflow goes further.
Measure the bottleneck without confusing the repositories
The authors repeat that Claude writes about 90% of Claude Code's own code. Claude Code has its own repository, separate from Anthropic's monorepo, whose growth from about 2,500 to over 30,000 daily merges was discussed in the earlier Scion article. Those are different workloads. The review goal is less total human time, including repair work after a bad change, with no rise in reversions or incidents. The team tracks approval counts, comments and waiting time, including median and 95th-percentile approvals. Doubling merged pull requests and reducing the reversion share is a reported observation over June–September, not proof that absolute failures or every team's risk fell.
Move verification before the request for approval
Before opening a pull request, engineers give the agent something executable to verify: tests, a build or screenshots, then use a fresh-context subagent to inspect the diff. Short project instructions and code-owner review rules preserve conventions and demand file-and-line evidence. Every pull request gets an AI review; other agents verify findings before they are posted, and style-only noise is filtered. The research preview increased the share of internal pull requests receiving substantive comments from 16% to 54%, a coverage indicator rather than a bug-fix success rate. Public availability on Team and Enterprise does not make every internal automation an included public feature, nor does this article establish its price.
Human approval is a risk boundary, not a formality
Open bug or security findings block merging until fixed or resolved with a reason; a later fixing commit can clear a finding automatically. A classifier assigns risk from the change, and another checks whether an extra person is necessary. Trusted authors making low-risk changes can use the zero-human route after green checks; ordinary changes require one person and the highest-risk changes two. Outsiders to the trusted list need another person, and sensitive owner-designated paths or CI workflows retain human review. An agent never fills a required human slot. The person directing an agent remains responsible and can count as a reviewer. Approval applies to a specific commit and must be renewed after changes the system cannot vouch for.
Review must converge instead of feeding itself
One pull request went through more than 50 rounds overnight. Looking back, about 59% of repeatedly flagged code had existed since the first review, rather than being newly introduced. The team now asks authors to reconsider the approach after three rounds, narrows later reviews to new diffs, suppresses repeated findings and asks agents to fix a whole bug class. This is not automatic permission to merge after three attempts. Confirmed incident-causing changes and their fixes form a daily evaluation set. Comment outcomes and tests on known-clean requests help measure noisy or harmful suggestions. Nightly review of existing code addresses old bugs without endlessly expanding a new change's review.
Release controls complete the process
The merge queue reruns checks with the current main branch so individually passing changes do not combine into failure. Flags, employee-first exposure, a latest channel before stable, weekly on-call and quick rollback limit the consequences when review misses something. The team says automatic circuit breakers that tighten approval as incidents rise do not yet exist. People still decide direction, review risky architecture and read evidence; the team has no formal process guaranteeing everyone understands the whole codebase. Our interpretation is that reversibility and observable risk justify carefully limited automation. An organization without those release controls should not copy the lightest approval path merely because an AI reviewer is available.
Why it matters
This is a distinct workflow account, not a second announcement of Scion or a new Claude model. It offers practical questions for agent teams: which evidence catches each failure, which changes always need people, and how an escaped defect becomes a regression case. AISubHub has not deployed or benchmarked this internal process.
Limits and availability
The results are Anthropic's own observations from an unusual, agent-heavy environment. Fewer human approvals are a goal under revert and incident guardrails, not a universal safety guarantee. Public review findings do not approve a pull request, and product availability does not establish free use or new personal subscription rights.
Sources and evidence
Source published: Oct 15, 2026 · Source checked: Oct 20, 2026
AI-assisted editorial articles checked against linked primary sources. Vendor claims are attributed; inclusion is not endorsement.