Set up layered agent code reviews
A team had Claude Code write most code, then used multiple review layers to catch bugs before merge.
- What they did
- Anthropic’s Claude Code team changed its review flow so the coding agent first checked its own changes by running tests, builds, or screenshots. A separate fresh-context subagent then reviewed the diff, and every pull request also went through parallel bug-hunting agents, with additional agents voting on each issue they found. Security review and CI still blocked merges, while a classifier decided whether humans needed zero, one, or two approvals based on risk. After that, changes moved through a merge queue, feature flags, and a gradual rollout.
- How it went
- Between June and September, the number of merged pull requests more than doubled, and the share that got reverted dropped by more than 40%. A convergence change also reduced the average number of reviews needed to pass from 2.08 to 1.32. They still do not have circuit breakers that automatically raise the approval bar when risk increases.
- Tips
- Keep the reviewer from rereading the same proposal forever: this team caps repeat review rounds at three and only asks it to inspect the new diff. Use the agent for checks, but keep a human approval gate for anything that could merge, ship, or change access.
Try it yourself with Claude Code
Review the pull request on [branch name] in [repo]. First compare the changes with the goal described in [PR description or ticket link]. Then inspect the diff for bugs, missing tests, security problems, and style issues against [conventions doc or file]. If useful, run or suggest checks such as tests, builds, or screenshots, but do not modify code. Stop before any action that would merge, deploy, send, delete, or submit anything, and ask for approval if that step is needed. Finish with a severity-ranked issue list and a clear recommendation on whether this is safe to merge.
Fill in the [brackets]. Agents can make mistakes, so review anything before it is sent, paid or deleted.
Get 3 agent uses like this every morning
With the day's top AI news. Free, and one click to unsubscribe.
Related agent uses
Make task handoffs require a commit ID
A developer paired a YAML todo script with a Claude Code helper that fetched the next job and closed it only after a commit ID was provided.
Use a browser agent to find and fix UI bugs
A frontend developer uses a browser agent and a coding agent together to spot UI issues, trace them in Chrome, and patch the code.
Investigate why a service is getting killed
Run a read-only SSH check to find the cause of a process death and rank fixes by risk.
Use Claude as an async coding partner
They paired with Claude and Jules while coding locally, then copied changes back and forth instead of sharing a live editor.
Set up a separate sandbox QA pass
Codex turned an existing Claude login into a second reviewer that checked an AI system without changing the user’s credentials.
Let Claude work from GitHub code and pause before posting
A commenter gives Claude source code and a feature brief, then reviews the draft only when PR review begins.