Generative UI
Also known as: GenUI, agent-generated UI, dynamic agent interface, generative user interface
Most AI products today put a chat box in front of a model and call it an interface. Generative UI flips that: the agent decides what UI surface to render based on what it's doing. Ask it to pull sales data and it returns a chart. Ask it to book a meeting and it renders a calendar picker. Ask it to confirm a risky action and it shows a structured approval form. The interface is an output of the agent, not a container for it.
Practically, generative UI requires a protocol for the agent to describe interface components and a frontend that can render them. The AG-UI protocol, developed by the CopilotKit team, is one approach: it lets a LangChain (a popular agent framework) agent emit structured component descriptions that a React (a JavaScript UI library) frontend renders as interactive elements. Vercel's AI SDK has a similar streaming UI primitive. The spectrum runs from controlled GenUI, where the developer pre-defines which components are available, to open-ended GenUI, where the agent can compose novel layouts.
For product builders, generative UI matters because it makes agent outputs sticky and verifiable. A table the user can sort and filter is more useful than a paragraph describing the same data. A form pre-filled by the agent and confirmed by the user catches errors before they cause problems. As agent workflows get more complex, the interface becomes the trust layer, and generative UI is how that layer gets built without coding a static screen for every possible agent state.