Generative UX Ends Static Screens at Scale

Gus Iwanaga demoed generative UX at AI Engineer, and the pitch is simple: a plain prompt turns into live React UI through an orchestrator and MCP tools, a remote-prompt path that affects any enterprise SaaS using declarative UI protocols.

He leads zero-to-one products at commercetools, an API-first platform with 300 plus APIs, and he framed the problem as 40 years of static screens piling cognitive load onto users.

How generative UX attacks actually work

Picture handing a junior designer your component library and a vague brief, then letting them place everything without a grid.

In practice the orchestrator classifies intent, pulls data from first or third party tools or MCP servers, then broadcasts a UI spec that an LLM renders inside a sandboxed iframe or as native components.

Iwanaga drove the risk home with four prompts for the same Q1 sales report that returned different KPI cards, date ranges, and charts. He rejected all of them for production.

Why it matters and what still isn't fixed

For builders the stakes are clear. The component catalog plus a Zod schema is the contract between agent and UI, so every prop and slot attribute dictates what can render safely.

His team tames nondeterminism with an atomic design hierarchy from layout to slot to sub slot to component, backed by a curated template catalog that teaches the UX agent what good looks like.

Still, placement stays nondeterministic, copy drifts, and teams no longer design pixels. They curate schemas, rules, and synthetic queries, a people and process shift Iwanaga called his biggest challenge.

Determinism itself isn't solved, so teams have to lock design tokens, validate specs against schemas, and keep a human in the loop for any UI that touches revenue or identity.

Declarative protocols like AG-UI, JSON Render, and OpenUI promise adaptive interfaces, but without that curation they ship confusion faster.