An agentic product doesn't wait for a click. It reads, decides and acts, sometimes across systems the user never opens. That changes what the interface is for. In a conventional product it is a set of controls. Here it is the place where a person decides how much to let the software do.
Say what will happen, and let people undo it
Before the agent does anything that matters, tell the user in their own terms. "I'll email the three vendors whose contracts expire this quarter" is something a person can approve. "Processing" isn't.
Where an action can be undone, make undo easy to find. Where it can't, as with sending money or an email, say so and ask first.
Let people set how much the agent does alone
New users often want to approve every step. A month later they only want to hear about the ones above a threshold. A setting with three levels covers most cases:
- Ask first: the agent proposes and the person approves
- Act, then report: the agent goes ahead and says what it did
- Act quietly: only for low-stakes work the person understands well
Keep a readable log, and plan for mistakes
An activity log is useful to users here, not only to developers. Write each entry in plain language: what happened, why, and what it touched. Someone coming back after a weekend should be able to see what the agent did in a minute or two.
The agent will be wrong sometimes. Show uncertainty plainly, and make a correction one click.