The Discipline Taking Shape: Notes on Agentic UX

The piece is right. A new discipline is forming, and the people who learn its grammar early will shape what software feels like for the next decade. The four principles it names (transparency, control, status, recovery) are the load-bearing walls of any honest agent interface, and I would not argue with any of them. What I want to do is sharpen the picture a little, because the principles are clean and the work is not.

Here is what I think the piece underplays.

The promise of an agent is that you do not have to watch it. The challenge of designing for an agent is making sure you can. Those two sentences sit at right angles to each other, and every pattern in the emerging literature is a negotiation between them. The progress ledger is a lovely object, but it solves the easy half of the problem. It tells you what happened. It does not yet tell you what almost happened, what was nearly chosen and set aside, what assumption the agent made when the data was thin. We are getting good at narrating action. We have not learned to surface judgment.

Transparency into reasoning sounds noble, and most current implementations confuse it with disclosure. A list of steps is not reasoning. A chain of tool calls is not reasoning. Reasoning is the part where a system decides which of three plausible paths to walk down, and that decision is almost always invisible in the timeline. The next wave of agent design has to make the road not taken visible. Otherwise the user is left with a record of motion and no way to question direction.

User control gets the same shallow treatment. Most products give you a stop button and call it agency. Real control is the ability to redirect mid-flight without starting over, to amend the agent's understanding of the goal rather than its current action. A pause and a resume are not enough. The interface needs verbs the user does not yet have: revise the brief, narrow the scope, change the audience, drop a constraint. We are still designing the kill switch when we should be designing the steering wheel the article gestures toward but the field has not built.

The coworker metaphor is doing more work than it can carry. A coworker has memory, accountability, a shared history, and the social cost of being wrong. An agent has none of these by default. When Salesforce calls its product a Coworker, it is making a promise the underlying system cannot keep. The honest version of the metaphor is closer to a very fast new hire on their first day, every day. The interface has to compensate for what the agent will never have on its own: continuity, context, and the kind of trust that is earned across weeks of small correct decisions. Until products treat memory and reputation as first-class design surfaces, the coworker frame will keep collapsing into something flatter.

Status communication is the principle I would push hardest on. Silence reads as failure, yes. But constant narration reads as anxiety. The best interfaces I have used recently know when to speak up and when to hold back, and that judgment is itself a design problem. An agent that talks at you through every step is harder to trust than one that withholds narration when its confidence is high. The progress ledger should be collapsible by default for routine work and expanded by default when something novel is happening. Right now most products do the opposite, and the result is a stream of small interruptions that train the user to look away.

Error recovery, the last of the four, is where I am most hopeful. The bar is so low that any thoughtful pass at it will feel like progress. The thing to watch for is whether recovery becomes a moment of repair or a moment of education. The best failures teach the user something about what the agent can and cannot do. They calibrate expectation. A good error message in an agent interface is closer to a postmortem than an apology.

So yes, the discipline is real, and the time to study it is now. The Gartner number will probably be right, give or take a quarter. But I would add a fifth principle to the four the article names, and I would put it underneath the others as their foundation: design for the relationship, not the interaction. A single exchange with an agent is the smallest unit of the work. The unit that matters is the arc across many exchanges, the slow accumulation of trust or distrust, the user's growing or shrinking sense that this system understands what they are trying to do. Most current products are still designing one screen at a time. The interesting ones are starting to design a relationship that lasts.

The patterns are forming. The harder work, the work that will distinguish a discipline from a style, is just beginning.

SourcesIn response to "Agentic UX: A Design Discipline Takes Shape." Sources referenced in the original article:
[1] Fuselab Creative — Agent UX in 2026
[2] UXmatters — Agentic AI in UX Design
[3] Onething Design — 5 Agentic AI UX Patterns
[4] Salesforce Ben — Agentforce Coworker