Ask most AI assistants to plan a weekend trip and you get a wall of text. Dates, hotels and prices, buried in paragraphs you end up reading twice. A good human assistant would just hand you three options, a date picker, and a Book button.

That’s Generative UI. Instead of replying with text, the agent replies with UI: a comparison card, a pre-filled form, a confirmation before you pay. It describes that UI in A2UI, the open protocol behind Flutter’s GenUI SDK, and your app renders it as real native widgets. The model never writes code. It can only pick from a catalog of widgets you control, and every tap goes back to the agent so the conversation keeps moving.

We’ll cover:

- How the agent, your widget catalog, and the app talk to each other.
- What breaks in production: invalid output, slow streaming, state drifting out of sync, and agents doing things nobody approved.
- When Generative UI is the right call, and when a normal screen is still better