You are about to let an agent do something a customer will see. Before you flip it on, run these five checks. None of them is clever. All of them are the ones people skip and regret.

1. Can it be turned off in one move?

If stopping the agent needs a deploy, it is not shippable. You want one switch, owned by someone on shift, that halts it without taking anything else down.

2. Does every action leave a record?

When it acts, you should be able to read back what it did, when and on what input. If you cannot reconstruct a run, you cannot defend it.

3. What is the blast radius?

Write down the worst thing it can do on its worst day. If that sentence makes you uneasy, narrow what it can reach before launch, not after.

4. Who gets paged?

Name the person, not the team. An alert that goes to everyone goes to no one.

5. What does the customer see when it fails?

Failure is a certainty, not a risk. Decide now whether they see a graceful hand-off or a broken screen.

Run the five. Ship the one you can turn off, read back, bound, page on and fail gracefully. Everything else waits a week.