AI agents can reason, call tools, and coordinate work, but most web apps still trap them inside a chat box. That is fine for conversation, but weak for real workflows where users need progress, approvals, forms, dashboards, and safe ways to correct the agent before anything changes.

In this session, we will take the red pill and look past the chat interface. Using a live Next.js, React, and TypeScript demo, we will build a support-triage agent that reads ticket context, streams progress into the UI, asks for consent before taking action, and renders an interactive resolution dashboard instead of dumping everything into chat.

Along the way, we will compare emerging agent UI patterns: AG-UI for keeping agent state and apps in sync, A2UI-style declarative interfaces, and MCP Apps-style interactive tool UIs. The goal is not AI magic; it is a practical architecture for building agent interfaces users can understand, inspect, and trust.

Attendees will leave able to explain where agent UI protocols fit, compare chat-first and protocol-first patterns, implement streamed agent state in React, evaluate consent and fallback boundaries, and troubleshoot stale state or unsafe actions.