Smashing Magazine: Beyond generative — the rise of agentic AI and user-centric design
Victor Yocco, UX researcher at ServiceNow and author of the forthcoming book on designing agentic AI experiences, makes a distinction that matters for how designers plan their work. Generative AI responds to a single prompt and then resets. Agentic AI understands a goal, plans a sequence of steps to achieve it, executes those steps, and adapts when something goes wrong. This difference — from reactive completion to proactive agency — changes what designers need to study, test, and account for.
The standard UX toolkit was built around usability. Can users find what they need? Can they complete a task? Agentic systems operate before users are asked to do anything. They detect problems, generate plans, and take actions on behalf of people who may not have a clear picture of what the system is doing. Yocco’s central argument is that when AI systems act as proxies for human judgment, the design questions shift from usability to trust, consent, and accountability.
He describes four distinct modes of agent behavior: detection (surfacing problems for human review), planning (generating a strategy that a person approves before execution), direct execution (taking action within defined parameters), and full autonomy (acting without requiring human input). These modes are not a progression — a given system might move between them depending on context. But each mode carries a different accountability structure, and designers need to be explicit about which mode is in effect and what recourse users have.
The research methods that follow are different from standard usability testing. Yocco recommends longitudinal studies over a fixed point-in-time evaluation, because trust in an agentic system builds or erodes over repeated interactions rather than in a single session. He also highlights the importance of error state research: not just whether users can correct a mistake, but whether they understand what happened, why the system acted as it did, and what they can do differently next time.
On the interface side, the article argues for non-intrusive notifications and clearly defined user controls. A well-designed agentic interface does not announce its own cleverness; it provides the user with timely, relevant information and a straightforward path to intervene. Yocco calls this design for confidence — the goal is a user who acts decisively because they understand what the system is doing, not one who hesitates because they do not know when to trust it.
The piece is practical in its framing. It draws on Yocco’s research experience rather than theoretical scenarios and lists specific research methods (longitudinal interviews, wizard-of-oz testing for planned agents, failure scenario walkthroughs) that teams can apply to projects in progress. Useful for UX researchers and senior designers who are moving from designing single interactions to designing systems that take actions on their own.