Agentic UX: designing for AI agents that work in the background

Most AI features answer when asked. Agents keep working after people look away: they watch data, talk to customers and run optimisations on their own. Nobody is watching when the agent acts, so the interface has to set its limits beforehand and explain what happened afterwards. These patterns come from three products I designed: Missed Call Assistant, Vytrac and Zentra.

In Missed Call Assistant, the rules the AI always follows sit on one screen, in plain words people can edit, such as “Never give a price over WhatsApp” or “Always mention the 1-year warranty”. Setup also asks who the assistant should leave alone, so it never messages those numbers.

If the professional calls back before the follow-up goes out, it is cancelled. Any action a person takes should override work the agent hasn’t finished yet.

Urgent cases, such as a caller reporting a gas smell, are escalated rather than handled. Notifications follow the same rule: they go out when the assistant needs a person to take over, not for every message it sends.

Vytrac’s AI health engine monitors patients all the time. When something needs attention, it arrives as an alert card with the actions on it: contact the patient, send a questionnaire, assign or adjust the threshold. Those cards appear on the provider’s phone too, because alerts come in away from the desk.

In Zentra, the optimisation agent runs when a manager starts it, and its suggestions appear next to the tickets they concern. Each ticket keeps an activity timeline beside its AI risk assessment, so anyone can see what happened and why.

In short: visible limits, people override the agent, escalate instead of guessing, end every report with a next step, and keep a trail with undo. See AI product design.