The Agent UI Stack: Building Interfaces for AI Agents with MCP Apps, A2UI, and AG-UI

The chat interface is hitting a wall. Users don't want to type five messages to book a restaurant — they want a date picker. They don't want a wall of text describing their sales data — they want an interactive dashboard.
In 2025-2026, a new protocol stack has emerged to solve this: MCP Apps (Anthropic + OpenAI), A2UI (Google), and AG-UI (CopilotKit) — three complementary standards that let AI agents generate and stream rich user interfaces. For web developers, this is a major shift: your frontend skills are suddenly the most valuable layer in the AI stack.
In this hands-on workshop, you'll build working agent UIs using all three approaches:

MCP Apps — Build a Python MCP server that returns an interactive HTML widget rendered inside a chat client. Your first "agent app" in 30 minutes.
A2UI — Create an agent that outputs declarative JSON UI instead of text. Build a TypeScript/React renderer that maps A2UI components to native widgets. Same agent, different renderings.
AG-UI — Wire up bidirectional event streaming so your frontend shows real-time agent reasoning, tool execution, and state updates — not just a final answer.
Production patterns — Security (iframe sandboxing, trusted component catalogs), multi-provider model routing, and lessons from upgrading a live enterprise AI platform serving 5,000 daily users.

A pre-built GitHub repository with backend scaffolding and starter templates will be shared before the event. Each section includes exercises with starter code, so attendees who fall behind can jump to the next section cleanly.