The Evolution of Generative UI for Agentic Applications

The Spectrum of UI Generation

Agentic interfaces are currently evolving through three distinct levels of complexity regarding how they present data to users. Static UI remains the most common approach, where an agent acts as an orchestrator that passes data and props to a predefined set of React components. Tools like the AG UI protocol and the Goose auto-visualizer exemplify this, as they rely on a fixed library of components built by developers.

Declarative UI represents a middle ground, offering a balance between flexibility and consistency. Instead of passing raw props, the agent generates a descriptor in JSON or YAML format. A rendering engine then interprets this descriptor to map data to the existing design system. This approach is widely used in server-driven UI architectures, such as those seen at Netflix, and is currently supported by tools like Vercel's JSON render.

Generative UI moves beyond predefined components entirely. In this paradigm, the model generates raw HTML, CSS, and JavaScript on demand at runtime. This allows for highly imaginative and context-specific interfaces that are not constrained by a pre-existing component library. The primary challenge with this approach is trust and security, as executing arbitrary code generated by an LLM requires strict containment.

The Role of MCP Apps in UI Delivery

Model Context Protocol (MCP) apps serve as the ideal delivery mechanism for generative UI because they provide built-in sandboxing. By utilizing a double iframe architecture by default, MCP apps isolate third-party code, mitigating the risks associated with executing model-generated front-end assets. This security model is robust enough that even first-party features, such as the Anthropic visualizer, leverage the MCP protocol to manage interaction and state.

Toward Human-Agent Collaboration

Current agentic interfaces are largely in a transitional phase, similar to early television broadcasts that merely replicated radio formats. The industry is currently stuck in a 'radio era' of UI, where agents are primarily used to output static visualizations. The future of agent interaction lies in shared, collaborative canvases rather than one-way output. The Excalidraw MCP app provides a glimpse of this shift, as it enables a persistent, shared space where both the human and the agent can modify artifacts in real-time. This moves the interaction model from simple command-and-response to a collaborative partnership.