AI agents don’t just answer questions; they take actions. They fill in forms, send messages, change settings and make purchases. That makes them useful, and it changes the design problem: people are no longer judging an answer, they are supervising a worker.
Watching every step doesn’t work#
If an agent asks for approval at every small step, people either give up on it or start clicking “Approve” without reading, which is worse than no checks at all. If it never asks, people can’t catch mistakes until it is too late. Good supervision sits between the two.
Show the plan first#
Before acting, an agent should show what it intends to do, in plain steps. A plan is quick to scan, and it is the cheapest moment to spot a misunderstanding.
Before
“Working on it…” followed, minutes later, by “Done! I’ve updated your bookings.”
After
“Here’s my plan: 1. Find flights under $250 on Friday. 2. Hold the best two. 3. Ask you before paying.” with Go ahead and Change plan.
Pause at the steps that matter#
Choose checkpoints by risk, not by habit. Good places to pause are:
- Before spending money or committing to something.
- Before sending anything to other people.
- Before deleting, overwriting or changing access.
- When the agent is unsure, or has found something unexpected.
Everything else can happen quietly, as long as it is visible afterwards.
Make the work reviewable#
A clear record of what the agent did, in order, lets people check the work at their own pace. Each entry should say what happened, where, and link to the result.
Undo builds trust#
People trust systems that forgive mistakes. Where actions can be reversed, make undo obvious and immediate. Where they can’t, say so clearly before the checkpoint, not after.
Let trust grow#
People are rightly cautious with a new agent. As it proves reliable, they may want fewer interruptions. Let them choose, per type of action, whether to be asked, told afterwards, or not bothered at all.
A quick checklist#
- Does the agent show its plan before acting?
- Are checkpoints placed at risky steps only?
- Is there a readable record of everything it did?
- Can people undo, or at least see what can’t be undone?
- Can people adjust how often they are asked?