Designing Agent UX People Trust
Interface patterns for agent products: showing progress, explaining actions, asking for approval, recovering from mistakes, and setting expectations.
Two agents can be equally capable and get opposite reactions. One gets used every day. The other gets tried once and quietly abandoned. The difference is usually the interface. The abandoned agent showed a spinner for ninety seconds, then announced "Done!" with no way to check what it did. When it got something wrong, the only option was to start over. Users couldn't tell whether to trust it, so they didn't.
Trust in an agent isn't a feeling you can design in with friendly copy. It's earned by an interface that sets accurate expectations, shows real progress, explains what happened, asks before it matters, and makes mistakes cheap to fix. This lesson gives you concrete patterns and code for each stage of a delegated task: before, during, and after. It builds on the run events, execution records, and approval endpoints from earlier in this module, and it sets up the product questions in the next lesson.
Design for calibrated trust
The goal isn't maximum trust. It's calibrated trust: users rely on the agent exactly as much as it deserves, for each kind of task. Over-trust means people stop checking and miss the mistake that matters. Under-trust means they redo the agent's work by hand, and the product delivers nothing.
Every pattern in this lesson serves calibration. It either helps users see what the agent is good at, or helps them catch it when it isn't.
Start with the surface. Chat is the default for agents, but it's often the wrong one:
| Surface | Works well for | Example |
|---|---|---|
| Chat | Open-ended, exploratory, back-and-forth tasks | Research questions, drafting with feedback |
| Task inbox | Delegated background work that finishes later | "Reconcile this month's invoices"; results arrive for review |
| Inline in an existing tool | Small, frequent assists where the work already happens | Suggested replies in a helpdesk, proposed edits in a document |
| Ambient or scheduled | Recurring work nobody should have to ask for | A morning digest, overnight triage, with notifications only for exceptions |
Pick the surface that matches how long the task takes and how often it happens. A ten-minute task shouldn't hold a chat window hostage. It belongs in an inbox with a notification when it's ready.