TopolearnReviewing is now your main job; skipping it is how plausible-looking bugs ship.
When Agent edits files, Cursor shows the changes as diffs: the old lines in red and new lines in green, grouped per file. Treat them like a pull request from a fast but sometimes careless colleague.
A simple review routine:
- Scan the file list first. Did it touch files you didn't expect, such as config, lock files or tests it "fixed" by weakening them?
- Read each change. Does the logic match what you asked? Are error cases handled? Did it invent a function or library option that doesn't exist? Models can do that confidently.
- Run the checks. Ask Agent "run the tests, the type checker and the linter and show me the output", or run them yourself. A diff that looks right but fails tests is not done.
- Use git as your record.
git statusandgit diffshow exactly what changed on disk, including anything a terminal command modified, and committing in small steps gives you safe points to return to.
Cursor can also help review. Agent Review runs a dedicated AI review of your local changes: type /agent-review in the agent input, run it from the Source Control tab to compare everything against your main branch, or enable it to run automatically. It offers a quick, cheap pass and a deeper, slower one for risky code, and it reads repository review rules from BUGBOT.md files. Treat its findings as suggestions, not proof of correctness.
If a change is wrong, say what is wrong and let Agent fix it, or roll it back (next lesson) and retry with a sharper prompt.
Example
# after Agent finishes
git status # which files changed (including via terminal commands)
git diff # exact line changes
npm test && npx tsc --noEmit && npm run lint # your project's checks
# in the Agent input
/agent-review # AI review of local changes (quick or deep)
git add -p # stage only the hunks you've reviewed
git commit -m "Validate empty email on signup"Remember
- Check which files changed before reading line by line
- Run tests, type checks and linters; a good-looking diff isn't proof
- Agent Review (/agent-review) adds an AI second opinion; git is the record
Watch out for
- Accepting changes where the agent made a failing test pass by editing the test's expectations.