Review, verify & guardrails
Treat the hand-back like a PR. Set rules and approvals.
Review the diff like you wrote it
After this you can review agent output with the same rigor as a teammate's PR.
Agent drafts. You ship.
Review starts before you read the summary. Open the diff and compare the changed files to the task you gave Agent. A clean explanation can still hide a missed edge, an accidental formatting sweep or a change in a file you never meant to touch.
Treat agent output like a pull request: inspect the diff, run the check and decide what ships.
Treat every agent hand-back like a PR from a fast teammate:
- 1Read every changed hunk.
- 2Run the code and the tests yourself.
- 3Spot-check the edges the agent didn't mention.
- 4Accept, refine the prompt or revert.
Never merge an agent change you haven't run. Tests passing in the agent's own run is a signal, not a guarantee.
Takeaway. Read every hunk, run it and own the result.
Self-check
QThe agent says all tests pass and summarizes a clean change. What do you do before shipping?
Guardrails: rules and approvals
After this you can set the two guardrails that make agent work safe to trust.
Start with two guardrails:
Rules reduce repeated corrections. Approvals protect actions with cost: package installs, file deletion, network calls, secrets, database writes and deploys. You want the agent to move quickly on reversible edits and slow down where a mistaken command changes shared state.
- Project rules
- Persisted conventions the agent follows every run (style, structure, do/don't).
- Command approvals
- Which commands run automatically vs. wait for your click.
If you correct Agent the same way twice, make it a rule.
Takeaway. Rules teach conventions. Approvals hold risky actions.