Ask at the consequential part
Routine choices should not become a stream of approval requests.
Retrospective field note for 2026-05-25, reconstructed on 16 November 2026. The archive date is not a claim of publication at the time.
Research, draft, send, buy. Different verbs, different commitments.
An assistant should preserve the one the user actually chose. Researching a trip does not authorise booking it. Drafting a message does not authorise sending it. Equally, an authorised task should not stop every few seconds to ask about another harmless implementation detail.
That balance needs to be designed into the product. Handle routine choices inside the brief. Pause when the amount, recipient, access or commitment changes. Make the question specific enough that the person can answer it once.
Keep a short record of consequential actions: what was requested, where it went and what the service confirmed. If the result is uncertain, inspect before trying again. Repeating the action just to keep moving can turn a communication problem into a real one.
The agentic direction at I/O makes these distinctions increasingly relevant. The interface should let someone delegate with confidence and then get out of their way. Constant babysitting is a failure of the workflow, even when every individual question sounds polite.
Google presented its agentic AI direction at I/O on 19 May. Source.