Guide · Code review

How to review an AI coding agent's pull request: a 10-point checklist

By Shamaz Khan · 18 Nov 2026 · 6 min read

Agent-written pull requests are fast to produce and slow to trust. The code usually reads well, which is exactly why problems slip through. This checklist focuses on what agents get wrong, in the order worth checking.

The checklist

  1. Does the file list match the task? Start with the list of changed files, not the code. Anything you can't explain from the task description is the first thing to question.
  2. Were any tests edited, skipped or deleted? Look for loosened assertions, skip markers and regenerated snapshots.
  3. Did config change? Build settings, lint rules, CI files, environment templates, tsconfig and similar. Small edits here have wide effects.
  4. New or upgraded dependencies? Check they're real, maintained and necessary. Agents sometimes add a package to solve a one-line problem.
  5. Deleted code. Removed lines are easy to skim past. Ask why each deletion is safe, especially "unused" helpers and odd-looking conditions that may be bug fixes.
  6. Renames and moves. Refactors nobody asked for make the diff bigger and the review harder. Ask for them in a separate PR.
  7. Error handling. Agents often handle the happy path well and swallow errors elsewhere.
  8. Security-sensitive paths. Auth, payments, permissions and data access deserve a slower read.
  9. Do the behaviours that matter still work? Not "do tests pass", but "does sign-in, checkout, export still work". Run the checks that prove it.
  10. Does the description match the diff? Agent-written PR descriptions can describe the intent rather than what actually changed.

Automate the first three and the ninth

Checks 1, 2, 3 and 9 are mechanical, so a tool should do them on every PR, leaving people to judge the rest:

stet scope set "Add CSV export" --allow "src/export/**,tests/export/**"
stet guard     # files outside the scope or in protected paths (incl. tests and config)
stet check     # locked behaviours still pass
stet pr        # PR description written from the real diff, scope result and check results

On GitHub, the Stet Action runs guard and check against the pull request's merge base and writes the result to the job summary:

- uses: shamazkhan/stet-action@v1

Keep PRs small

The single best review aid is a small PR. One task, one scope, one PR. If an agent's change can't be reviewed in ten minutes, split the task before asking for more code.

FAQ

How do I review AI-generated code faster?

Review the file list against the task first, automate the mechanical checks (out-of-scope files, test and config changes, key behaviours), and keep each PR to one small task.

Can AI review AI-generated pull requests?

AI reviewers are useful for spotting bugs and style issues, but they judge the code they see. Pair them with deterministic checks on scope and behaviour that do not depend on a model.

What should I look for first in an agent's PR?

Files that have nothing to do with the task, and any change to tests or config. Those are where agent PRs most often go wrong.

Put guardrails on your AI coding agent

Stet is a free, open-source CLI and MCP server for Claude Code, Cursor, Codex and Copilot. It runs locally: no account, your code never leaves your machine.

npx stet-cli init · Install · Setup guide · GitHub

Related