Designing for the agent that acts, not the one that answers
By Allan Leone on
Gartner expects 40 percent of enterprise apps to carry task-specific agents by the end of this year, up from under 5 percent last year. Chat patterns do not cover any of the hard parts.
The projection going around is that 40 percent of enterprise applications will embed task-specific agents by the end of 2026, against under 5 percent in 2025. Treat the exact figure with the usual caution. The direction is not really in doubt, and the design consequence is being underestimated.
An agent that answers is a conversation. An agent that acts is a delegation, and delegation has an interface literature that chat borrowed none of.
The four questions chat does not answer
- What is it doing right now, at a glance, without reading a transcript.
- How do I stop it, and what state am I left in if I do.
- What did it change, expressed as a diff rather than a narrative.
- What happens when it half finished, which is the normal case, not the edge case.
A streaming text box addresses none of these. It shows you prose about work while the work is happening somewhere you cannot see.
Progress is not the same as status
A spinner says something is happening. Status says which step, of how many, on which object, reversibly or not. The difference matters most exactly when the user is nervous, which is exactly when they are watching.
The pattern that has worked for us is a plan the user can see before it runs, and the same plan annotated as it executes. It borrows from install wizards and deployment pipelines, both of which solved this problem long before anyone said agent.
Undo beats confirm
Confirmation dialogs before every action destroy the reason to delegate. If I have to approve nine steps I have not saved anything, and by step four I am clicking without reading, which is worse than not asking.
Ask once, at the boundary that actually matters, usually the irreversible one. Then make everything inside that boundary undoable and show the user what changed. Cheap reversal is worth more than expensive permission.
Error recovery is the whole product
The demo shows the happy path. The product is the other one. An agent that fails halfway through a six-step task has left the user in a state they did not create and cannot see, and the interface owes them an accounting of it.
We now design the partial-failure screen before the success screen. It is a bleak way to start and it produces a better product, because the success screen is easy and everybody already knows what it looks like.
Tags: ux, ai, product