Agent Experience (AX): The New Design Discipline

Forty-two years ago, Don Norman coined the term “User Experience” (UX) while at Apple. That idea launched decades of innovation in interface design, mobile apps, and digital products.

With the rise of agentic architectures, we’re entering a different era with a new participant and new interface needs. This shift demands a new discipline around Agentic Experience, or AX.

When the User Isn’t Human

One trend is unmistakable: we are moving toward agentic architectures. Work that required human decision-making and coordination is being enhanced by agents that can reason about context, plan multi-step actions, and execute across tools.

AI agents are now becoming critical system participants, and they operate differently than humans. They consume information, interpret context, make decisions, and take action in ways that humans do not. That fundamentally changes interface design needs.

Traditional UX optimizes how a human uses a tool. It requires deep understanding of buttons, menus, visual hierarchy, screen layouts, and feedback loops, all orchestrated to make interaction intuitive and efficient for humans.

The challenge is no longer only about creating better screens. It’s now about structuring information and exposing capabilities so agents can understand and act deterministically.

Agents don’t scan visual hierarchies or rely on spatial memory. They parse data, weight probabilistic relationships, and follow explicit instructions. Those differences aren’t edge cases; they’re a new, core design constraint.

What AX Actually Means

UX reduces friction for humans to work in ways native to human cognition. We’ll still need that, but we now need AX to reduce friction for agents to accomplish tasks, in ways native to LLMs.

This shift forces us to rethink how we present our system interfaces. Good AX means several things:

1. Information must be structured with an explicit schema. Agents need clear data types, defined relationships, and unambiguous documentation. What a human might intuit from context, an agent needs spelled out.

   
   

2. Capabilities must be exposed granularly. Where a UI might offer a “Schedule Meeting” button that opens a form, an agent needs to access discrete actions such as check availability, propose times, send invitations, and handle conflicts. Each action should do one thing precisely and return structured feedback.

   
   

3. Different LLMs process information differently. Claude, GPT, and Gemini each have distinct architectures and training dynamics. Your system’s documentation and data formats need to work across models, not just today’s leading LLM.

   
   

4. LLM-native formats drive better results. Markdown for documentation. TypeScript or JSON Schema for data structures. XML when hierarchical relationships matter. These aren’t just convenient. They align with how LLMs were trained to parse and generate information.

   
   

Why This Matters Now

We spent thirty years designing for humans. The next decade will be about designing systems where agents and humans collaborate as peers. The interfaces we build won’t just live on screens; they’ll exist as programmatic capabilities that agents invoke, chain together, and reason about autonomously.

Some will see this as a technical problem: better APIs, more documentation, cleaner schemas. That’s necessary but insufficient. AX is a new design discipline. It requires understanding how agents process information, anticipating failure modes in reasoning, and building systems that degrade gracefully when agents misunderstand or lack context.

The companies that master this first will have a multi-year competitive advantage. They’ll be truly AI-native while others are still retrofitting chatbots onto legacy workflows.

AX isn’t the next chapter of UX. It’s the prerequisite for competing in the AI age.

Learn More About the Author