Human-centered design took recognizable form in the 1980s as a counterweight to technology-first thinking (Norman and Draper, 1986). Its principle was plain: build around how people live and think, not around what engineers find convenient.

Forty years later, the interface has started acting on its own. It reads context, pursues a goal, and takes steps before a person can inspect each one. The human is still said to be at the center. The center, however, has begun to move.

The Inversion Problem

Traditional interfaces waited. Every meaningful action required a gesture: a click, a tap, a command. That pause between intention and execution gave people a small but useful form of control.

Agents remove much of that pause. Given a goal, a system can execute hundreds of steps without presenting each decision to a person. Automation has always varied by function and degree; agentic systems make that continuum easier to traverse and harder to see (Parasuraman et al., 2000). This makes agents useful. It also makes a ceremonial human-in-the-loop remarkably easy to manufacture.

The design problem has changed. We are no longer arranging controls for every action. We are deciding when a machine may act, when it must ask, what a person can inspect, and whether “stop” still means stop.

Principles for Agentic UX

Five working principles follow. They extend established human-AI interaction guidance on visibility, correction, control, and failure, along with risk-management guidance that treats trustworthiness as a property of the whole sociotechnical system (Amershi et al., 2019; NIST AI RMF 1.0, 2023).

1. Make action visible. People need to see what an agent did, why it did it, and what changed. Show the work.

2. Interrupt where judgment matters. Requiring approval for every step defeats delegation. Requiring it for none defeats oversight. Identify decisions that are costly, irreversible, or morally significant, and stop there.

3. Preserve human override. Give people a clear way to pause, redirect, or cancel. An override buried in a settings page is an archaeological artifact, not a control.

4. Make failure recoverable. Agents will err. Record what happened, explain the failure, and support undo where undo remains possible.

5. Make autonomy earn its keep. Begin with narrow authority. Expand it when evidence justifies the expansion, not when a product roadmap would find it convenient.

These are hypotheses, not commandments. Systems will test them. People will bear the cost when they fail.

Sources