Part 10 of 12 · 1 min read
Quick notes
- The interface decides whether people trust and use the agent. Every screen must answer: what is it doing now, what needs you, and how do you stop or undo it?
- Show progress in layers, not as a log: a status line, a never-collapsed "needs you" area, a grouped timeline, and a changes-and-controls area.
- Approvals: tier them, batch them, default to no. Fewer, better-placed prompts protect the dangerous ones from approval fatigue; an unanswered request expires as denied.
- Undo is the best safety feature: an action log with inverse operations, explicit irreversible actions, and a delayed commit for outbound messages. Offer pause, resume, cancel and smart resume.
- Say uncertainty when it is calibrated, ask when it matters, fail gracefully, and show provenance and cost. Then test the interface with a few real users.