A working primer on agentic UX

AI agents can complete real multi-step work now. The design challenge has moved: it is no longer making the agent capable, it is making its work legible.

Three patterns carry most of the weight.

Scope preview. Before anything runs, show what the agent will do, what it can touch, and what it will not do. Delegation starts with a clear contract.

Progress that reads like a colleague’s status update. Which step is active, what has been produced so far, what is queued. Not a percentage bar over a black box.

Confidence and undo. Agents will make mistakes. Products that admit this, with cited sources and one-click reversal, keep trust through errors. Products that present every output with equal certainty lose trust on the first visible failure.

I keep a growing library of these patterns in my AI design system work. The interesting part is how boring the winning patterns are: they are the same trust mechanics good teams use with each other.