How the Claude Code team does code review

TL;DR Highlight

Moving most review to AI doubled merged PRs while reverts fell over 40%.

Who Should Read

A backend or platform developer whose team's human review is bottlenecked by the flood of agent-written PRs

Core Mechanics

PRs doubled, reverts down 40%
- From June to September, merged PRs to Claude Code's codebase more than doubled.
- Over the same period, the share that had to be reverted fell by more than 40 percent.
Review moved from a final gate to continuous checks
- Before, you wrote code, opened a PR, and waited for a colleague.
- Now review runs like spellcheck, and Claude reviews and fixes most changes at least once before a human looks.
Agents find bugs, other agents vote
- Several agents hunt for bugs at the same time.
- A separate group votes on each finding, and it is posted only if enough agree it's real.
Human approvals scale by risk: 0, 1 or 2
- A classifier reads the change and sets the risk tier.
- Docs and copy need no approval once automated checks are green, caching, auto-updater and system prompts need one person, and the highest risk needs two.
- An agent's approval never counts as a person's.

Community Reactions

Opinion that this only works at a lab
- Some argued Anthropic can afford this with virtually unlimited tokens, while other teams should keep reading code.
- One commenter recalled smaller teams getting hurt copying Google's process.
Compliance and security concerns
- One asked how this passes SOC2, since their auditor required at least one other human to review code before production.
- One pointed out the README of the security review action says it is not hardened against prompt injection.
Doubts that flag-gated changes are low risk
- One said the risk materializes when the flag is turned on, and asked whether review happens then.
- Another argued that asking Claude for explainers is search, not real knowledge.
Experience: compiled projects make this hard
- One commenter said their toolchain is about 200GB and integration tests take hours, so such agent feedback loops are impractical.

How to Apply

Build a reviewer eval set from last month's reverts
- In the next 30 minutes, list incident and revert PRs with where the bug lived.
- Run your current AI review prompt on them and count how many it catches.
Write a REVIEW.md telling the reviewer what failure looks like
- State what failure looks like, which paths to skip, and a cap on nits.
- Require a file and line as evidence for each finding, and start the summary with a count.
Add a round cap and diff-only re-review to AI review
- After three rounds, make the reviewer ask the author to reconsider the approach.
- From round two, review only the changed diff and forbid repeating posted findings.
Create a risk tier table for changes
- Always require a human for paths like billing, the auto-updater and CI workflow files.
- Waive human approval only for low-risk paths such as docs and additive changes behind a flag.

Try It

copy & run

# Paste into Claude or ChatGPT (simplified version of the article's method: agent voting + convergence rules)

You are a code reviewer. Find bugs using only the diff below.
Rules:
1. For each finding, give file and line as evidence. Do not guess from names.
2. Exclude style, cleanup, and diagnostic-only findings.
3. Tag each finding Important / Nit / Pre-existing. A bug this PR did not introduce is Pre-existing.
4. Group the same class of bug and list every location at once.
5. Start the summary with the count of findings, and say plainly if there is nothing to fix.
6. At the end, try to refute each finding yourself and keep only those that remain real bugs.

PR intent: {{PR description}}
Diff:
{{diff}}

Next step: after fixes, add "do not repeat earlier findings, review only the changed diff" to the same prompt and run round 2.

Terminology

RevertUndoing a merged change. It is the quickest way to recover from a bad deploy, so it is used as a quality metric.

Merge queueA queue that re-runs all checks against the latest main with your change applied. It stops two PRs that pass alone from breaking main together.

Feature flagA switch that turns a feature on or off without a deploy. You can disable a problem without reverting code.

Fix-forwardFixing a problem by pushing a new commit instead of reverting.

REVIEW.mdA file written by code owners to tell the reviewer what matters in their part of the code.