Manage agents by proof.

AI agents are easy to start and hard to supervise. The mistake is judging the work by effort, tool logs, or confident summaries. Proof-based management judges the work by what can be checked.

1. Name the outside-world result. "Update the page" is weak. "The public pricing page shows the new refund line" is checkable.
2. Define what must not change. Good boundaries prevent an agent from solving one problem by breaking a quiet dependency.
3. Ask for the actual evidence. A screenshot, public URL, exact receipt, exported file, or Stripe transaction is stronger than "tests passed."
4. Require an unproven list. If the agent cannot name what it did not verify, it probably does not know where the risk is.

The $1 version.

The Agent Handoff Checklist turns this method into a reusable handoff and review sheet you can paste into agent work before the mess starts.