Set how much it can do
NewHow much should the assistant do on its own, and who decides?
Start careful, let people raise it, confirm only what cannot be undone, and keep the rest visible and undoable.
Try it
Send an email
Send a reminder to a customer.
ana@example.comThis command requires approval
Limits for this run
Spend
120 of 200 euros used
A short account of the result and anything that needs you.
Sent 3 reminders and paid 1 invoice.
How to design it
- 1
Start careful
Begin with the assistant asking first. People can see what it does before they trust it.
- 2
Let people raise it
Offer a few clear levels and a limit on time or money. Let people change them at any moment.
- 3
Confirm only what cannot be undone
Ask about payments, deletes and messages. Let the rest happen, in plain view, with Undo.
- 4
Show who approved
Mark each action as approved by the person, by a saved rule, or by nobody.
Do and avoid
Do
- Say in plain words what each level lets the assistant do.
- Say exactly what "allow for this session" covers.
- Show the limit and how much is used.
- Let people choose how much detail they see.
Avoid
- Do not ask about every step.
- Do not start on full freedom.
- Do not hide who said yes.
- Do not make the setting hard to find or change.
Why it works
Anthropic reports that people approved about 93% of permission prompts, and that attention falls as prompts pile up. It also found that new users let the agent run freely in about 20% of sessions, against over 40% for experienced users. So the sources point to asking only about what cannot be undone, and making everything else visible and easy to reverse.