Lesson 5 of 6
Agent UX (approve / observe / undo)
7 min read
An agent that can book, buy, send, and delete on your behalf is useful right up until the moment it does the wrong one. Then it's a lawsuit.
Trust needs a leash
Give an [agent](glossary://agent) real actions and the UX question becomes: how does a person stay in control? Three affordances do most of the work — approve risky steps before they run, observe what the agent is doing, and undo what it did. Run the plan in the scene: safe, reversible steps go on their own; the risky, irreversible one stops and waits for your yes.
Automate the reversible; gate the irreversible. The human in the loop approves exactly the steps that can't be taken back.
Approve, observe, undo
Design all three. Approve: put a checkpoint before anything costly or permanent — pay, send, delete. Observe: show the plan and a live activity log so the agent is never a black box. Undo: make actions reversible where you can, and a clear stop-and-rollback where you can't. The more an agent can do, the louder these three must be.
The autonomy an agent earns is bounded by how well the user can approve, watch, and reverse it.
Not every step deserves a prompt — gating the safe ones too trains users to click 'Approve' without reading. Reserve the interruption for the steps that actually matter, so it still means something when it appears.
The shape of it
- —Gate risky, irreversible actions behind explicit approval; auto-run safe ones.
- —Make the agent observable — show its plan and a live activity log.
- —Provide undo, or a clear stop-and-rollback when undo isn't possible.
Your agent can draft emails and also send them and issue refunds. Which step needs an approval gate?
Continue in the app
Take the whole Designing AI UX (for uncertainty) course — tracked
Get your personalized path, progress and streaks in the app — this lesson and every next one, in order.