From UX to AX

The core shift

From the gulf of execution to the gulf of evaluation

For forty years, interface design lived in what Don Norman called the gulf of execution: the distance between what you want and figuring out how to make the software do it. Menus, buttons and flows existed to close that gulf. When an agent does the doing, that gulf mostly disappears — and a different one opens up. The hard problem becomes the gulf of evaluation: how do you know the agent did the right thing, well?

This is the through-line in the most credible 2026 thinking. John Maeda’s Design in Tech Report 2026, subtitled “From UX to AX,” frames the move as going “from crafting interfaces to orchestrating outcomes,” and describes an agent as “a loop: action, outcome, feedback, correction.” confirmed Nielsen Norman Group frames the same idea as outcome-oriented design — the designer specifies goals and constraints, and gives up some control to the system, rather than hand-placing every screen. likely

What’s solid

A named discipline, and a kit of trust patterns

“Agent Experience” (AX) was coined by Netlify’s CEO Mathias Biilmann in January 2025, defined as “the holistic experience AI agents will have as the user of a product.” His own framing is technical — four pillars of Access, Context, Tools and Orchestration: can an agent reach your product, understand it, act in it, and combine those actions? confirmed vendor-origin

On the human side, a small set of trust patterns keeps recurring across the practitioner writing — and Smashing Magazine’s February 2026 piece covers the same ground as “control, consent and accountability.” The recurring kit:

• Intent Preview — the agent states its plan and you approve, edit or reject it before it acts. The single most important pattern: it prevents the “what just happened?” moment that destroys trust.
• Autonomy Dial — a slider from “full manual” to “full autonomy,” set per task and context.
• Confidence Signal — the agent shows how sure it is, so you know when to look closer.
• Explainable Rationale & Action Audit — show the work, and keep a reviewable trail of what changed.
• Escalation Pathway — a clear, quick route to a human when it matters.

The exact six-name taxonomy is really one careful synthesis rather than a settled standard — so treat the list as a useful scaffold and the theme (preview, control, confidence, audit, escalation) as genuine consensus. single-source taxonomy

Underneath the patterns sits a discipline borrowed from safety engineering: human-in-the-loop, calibrated by stakes. You don’t gate every action — only the consequential, hard-to-reverse ones — and you grant agents more autonomy as they earn it (progressive autonomy). The maxim doing the rounds: “policy before prompt” — validate what the agent proposes as if it came from an untrusted source, rather than trusting the model to behave. confirmed

The other half of the story is generative UI: agents assembling the interface on the fly instead of you hand-building every screen. Gartner expects 30% of new apps to use AI-driven adaptive interfaces by 2026 (up from under 5%), and 40% of enterprise apps to embed task-specific AI agents by 2026 (up from under 5% in 2025). likely — forecasts And trust is now the design problem: NN/G’s State of UX 2026 warns that as agents ship before they’re ready, “AI slop” turns the sparkle into noise — and the backlash is growing. likely

The contested zone

Is the interface really dying?

The loudest claim comes from Jakob Nielsen, who declared that “graphical interface design will disappear as we know it,” with human interfaces becoming obsolete as we shift to intent-based interaction. The field largely rejected it — Adrian Roselli’s “Jakob has jumped the shark” is representative — and the practitioner consensus runs the other way: higher stakes mean more human verification, not less. contested

The counter-case is well argued. “Chat is a terrible interface for agents”; the “chatbot-first” mindset “fails at scale”; you don’t chat with a deployment pipeline — you expect it to handle things and tell you when something matters. Graphical interfaces still win decisively for structured, repetitive and high-risk actions, where they minimise errors and cognitive load. The honest 2026 position is therefore hybrid: language for intent, interface for control. confirmed

Two more honest caveats. First, AX is partly a vendor narrative — the term is Netlify-coined and much of the supporting writing is agency SEO. It’s real (Bessemer made it a “law,” Sequoia talks of “agent-led growth,” independent voices like Maeda and NN/G take it seriously) but the certainty is inflated. Second, accessibility is a live fight: Nielsen’s claim that agents make accessibility “irrelevant” by translating for disabled users was widely called tone-deaf — disabled people want to use technology, not be spoken for. contested

How deep, and how honest