Most AI answers arrive as a paragraph. This one arrives as a UI — a table, a chart, a row of stats — assembled on the spot, matched to whatever you actually asked. No developer pre-built a screen for "show me overdue invoices"; the model decided that question needs a list, this one needs a bar chart, and rendered accordingly. This is the shape of A2UI (Google's generative-UI protocol, open-sourced 2026) — an agent that returns structure, not just text.
CHECKING ENGINE…SudsCo Ops Desk · fictional data
Ask the ops desk · —
Ask a question above — watch what renders.
Honest label: on a Prompt-API-capable Chrome, Gemini Nano generates a real, schema-constrained UI spec live — badge reads NANO LIVE. Elsewhere, the four sample questions replay pre-computed specs (labeled SIMULATED) so the interaction is honest either way; typing your own question needs the live engine.
Every screen your team asks for — without building every screen.
Dashboards break the moment someone asks a question nobody designed a view for. An agent that renders structure instead of text doesn't have that ceiling — it builds the view the question actually needs, every time.