UX Collective: AX is just the orchestration layer

Adrian Levy’s article in UX Collective opens with a simple observation: “The work didn’t vanish with the screen. It fell one layer down.” That single sentence frames the argument he spends the rest of the piece developing — that as AI agents take over more of the direct interaction with users, design work doesn’t disappear but moves to a different location in the system.

Levy picks up from John Maeda’s framing of the shift from UX (user experience) to AX (agentic experience) and tries to locate exactly where AX lives. His answer is the orchestration layer: the part of a product where intent is received, agents are directed, and decisions about what happens next are made. This layer sits between what users say they want and what AI systems actually do. Designing it means deciding how ambiguous instructions get clarified, how competing agent actions get prioritized, how errors surface in ways that preserve trust, and how the system knows when to hand control back to the human.

The distinction Levy is drawing is not merely semantic. In a conventional product, the interface is where most design decisions live — visual hierarchy, interaction patterns, copy, state handling. In an agentic product, much of what the user experiences is not a designed interface at all. It is a sequence of AI decisions unfolding in response to a broad directive. The designer’s job is not to lay out that sequence in advance but to define the principles, guardrails, and resolution paths that govern it when things go as expected and especially when they don’t.

Levy is careful not to overstate the case. Screens and interfaces still exist, and the craft of designing them still matters. What has changed is their relative importance. The orchestration layer — invisible, structural, logic-rather-than-aesthetic — increasingly shapes whether a product feels coherent or chaotic.

The practical implication for designers is that skills in systems thinking, error state design, trust signaling, and clear delegation logic are becoming more central than they have been. Writing clear task definitions for agents, thinking through edge cases in multi-step flows, and deciding when to surface uncertainty to users rather than resolving it silently are all design problems — just not the kind that live in a component library or a Figma file.

The article is worth reading alongside documentation on building agentic products, where the same questions appear from the engineering side. Levy’s contribution is to name the designer’s stake in them clearly: orchestration is not a byproduct of the system architecture. It is where design is actually happening now.