Contents

The Claude Code team is shifting code review from a process where humans read every line to a system that combines automated checks with human judgment based on risk. The goal is to reduce review time without increasing reverts or incidents.

Key Points#

  • Most changes are checked with Claude and automated checks before a PR is opened.
  • The number of human approvals depends on risk; low-risk changes may pass with automated checks alone.
  • Separate groups of agents verify AI review findings, and authors address each finding.
  • Reviews are capped at three rounds, after which the focus shifts to new diffs.
  • Revert and incident rates serve as guardrails when reducing human review time.

Use Human Time as the Measure#

In Anthropic’s monorepo, the number of changes merged per day grew from roughly 2,500 in May 2026 to more than 30,000 in September. Claude Code has a separate repository, but its volume of changes is also growing, making it difficult for humans to read everything line by line.

The team focuses on “what could happen because of this change, and which checks would catch it,” rather than “whether a human read the code.” Human time includes more than time spent reviewing. If fixing bad changes consumes that time, shorter reviews offer little benefit.

The team therefore measures human approvals and comments per PR, along with time spent waiting for someone to act, while treating avoidance of increased reverts and incidents as a constraint. The article reports that from June to September 2026, merged PRs for Claude Code more than doubled while the proportion reverted fell by over 40%.

Place Checks Before and After Opening a PR#

The first checks happen before a PR is opened. The team provides ways for Claude to run tests, builds, visual comparisons, and other checks and show the results. Some team members also use subagents with fresh context to examine diffs for bugs, simplify code, and verify behavior.

Every PR undergoes a Claude review and a security review, with separate groups of agents validating the findings. Style preferences and poorly supported findings are filtered out, and the remaining findings are posted with severity ratings. Authors must either fix them or resolve the threads with an explanation. PRs cannot merge while findings remain unresolved.

Linters, type checks, platform-specific tests, and builds then run. The merge queue reruns checks with the changes applied to main. This prevents cases where two PRs pass individually but break when combined.

Adjust Approval Requirements by Risk#

The changes are assessed for risk to determine how many human reviewers are needed. The criteria shown in the article require no approvals for low-risk changes, one for medium-risk changes, and two for high-risk changes. Wording fixes and new features behind flags are examples of low-risk changes; regular feature changes, caching, and update mechanisms are medium-risk examples. High-risk code and designated areas that always require two reviewers fall into the high-risk category.

Reduced approval requirements apply only to work by trusted repository members. Designated paths, such as CI configuration, require human review, and agent approvals do not substitute for human approvals. New commits require fresh approvals. The person who asks an agent to do the work also remains responsible for the change.

After merging, flags and gradual rollouts limit the impact. Claude Code rolls out internally first, and public releases reach the latest channel before the stable channel. If problems arise, the team disables the flag or reverts the change. More automated checks cannot prevent every mistake in advance.

Keep Reviews from Dragging On#

When AI repeatedly reviews AI-written code, fixes and rechecks can become endless. The article describes a CLI PR that went through more than 50 rounds of review and revision overnight. These rounds did not always uncover new bugs; in some cases, they simply flagged existing code one piece at a time.

The team caps routine reviews at three rounds, then revisits the implementation approach if more are needed. After the initial checks, reviews focus on new diffs and encourage fixing similar issues together instead of one location at a time. They also avoid repeating previously posted findings.

Treating the number of findings alone as success also increases false positives. The team tracks how people respond to findings and checks whether reviews prompt unnecessary fixes to otherwise sound PRs. Past incidents and reverts are added to an evaluation set that reproduces them, and checks run daily after changes to the review system.

Decisions Humans Still Make#

Humans remain responsible for deciding what to build, how to design it, and whether high-risk changes should ship. People review billing, update mechanisms, and CI configuration, while experienced reviewers examine design patterns that other changes will also use.

Reviewers examine test results, commands run and their output, and visual evidence. If they cannot identify what problems the checks might miss, they request a detailed review instead of giving a quick approval. Authors of small changes are also responsible for fixing any problems they cause.

Reducing review workload also requires keeping changes small and making them easy to roll back if they fail. The team investigates which checks or procedures missed a problem without turning incidents into personal blame. Whether human review can be reduced depends on these operational practices too.

Next Steps#

Start by choosing one repository and adding tests, risk, and rollback plans to its PR template. Then classify changes as low, medium, or high risk and decide how many human approvals each level requires. Track both human review time and reverts or incidents, then revisit the criteria after a few weeks.

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

Contents