# Anthropic outlines Claude Code’s AI-led review process

> The team aims to merge most changes with at most one human reviewer, while acknowledging unresolved risks and missing safeguards.

Oossa · 2026-10-29 · https://oossa.com/en/anthropic-outlines-claude-code-s-ai-led-review-process

Anthropic says Claude Code now writes about 90% of its own code, and the team is changing how it reviews that work. In a post by Addy Osmani, Morgan Westlee Lunt and Qing Wang, the team describes a goal: most pull requests should merge with at most one human involved, who may be the person who wrote the change, without increasing reverts or incidents.

The team says reviewers do not need to hold a complete mental model of Claude Code. Instead, they ask Claude to explain the part of the codebase they need to understand. That approach drew criticism in a Hacker News discussion: commenters questioned whether relying on AI explanations leaves anyone able to debug a failure when the model’s understanding is wrong.

## How the review gates work

A classifier assigns each change a risk tier based on the change itself, not its author. Low-risk changes include documentation, copy and additions behind a feature flag; they can pass without human approval. Medium-risk changes, including work on caching, the auto-updater and system prompts, require one person. An agent’s approval does not count as a human approval, but the person who prompted a change can count as the required reviewer.

The review process has its own failure mode: agents act on feedback that a person might ignore. Anthropic describes one pull request that went through more than 50 rounds of review and fixes overnight before the author stopped it. The team says it capped the number of rounds after finding that simply running more reviews made things worse. It also found that about 59% of the code flagged in later rounds was already present when the reviewer first looked.

## What remains uncertain

A separate example from Claude Code’s public repository shows the limits of generated explanations. In August, a Claude-generated response said issues were closed for inactivity when they lacked information needed to reproduce a problem. The repository’s rules, however, allow issues to close after 14 days marked stale and another 14 days, with an exception for issues with at least 10 upvotes. A developer counted more than 6,000 issues labelled as having reproduction steps that were still closed as inactive. The response’s suggestion to comment to reopen an issue also ran into a workflow that locks closed issues after seven days.

Anthropic says it does not yet have circuit breakers that tighten approval requirements if its review metrics worsen. Its Code Review product is generally available on Team and Enterprise plans. The security-review GitHub Action is open source, but its README warns that it is not hardened against prompt injection and should only review trusted pull requests.

## The facts

- Anthropic says Claude writes about 90% of Claude Code.
- The team reported one pull request that went through more than 50 review-and-fix rounds overnight.
- Anthropic says about 59% of code flagged in later review rounds was already present at the reviewer’s first pass.
- Code Review is generally available on Anthropic’s Team and Enterprise plans.

## Why it matters

For teams considering AI-led code review, Anthropic’s process shows that low-risk changes may merge without human approval, while medium-risk changes still need a person. The company says it has not built circuit breakers to tighten approvals when review metrics worsen, so the post does not establish how the process would respond if its safeguards began failing.

## Sources & references

1. [Anthropic Says Nobody Needs a Mental Model of Claude Code. Practitioners Are Testing That](https://www.infoq.com/news/2026/10/anthropic-code-review/?utm_campaign=infoq_content&utm_source=infoq&utm_medium=feed&utm_term=AI%2C+ML+%26+Data+Engineering) – InfoQ, 2026-10-29

Last updated: 2026-10-29