Over the past two years the agent stack has standardized its edges one seam at a time: MCP for tools and data, A2A for agent-to-agent handoffs. The last seam — how an agent streams what it is doing into a user interface — is now visibly consolidating around AG-UI.

AG-UI, the Agent-User Interaction Protocol, describes itself as “an open, lightweight, event-based protocol that standardizes how AI agents connect to user-facing applications.” Introduced by the CopilotKit team and released under the MIT licence, it works by streaming structured events — messages, tool calls and state updates — from an agent runtime to a front end.

Adoption is no longer theoretical. The protocol is documented as an integration in pydantic-ai; it has been proposed for Spring AI; Google's Agent Development Kit now carries an AG-UI middleware; and the Chainlit project tracks AG-UI support alongside the newer A2UI. On GitHub, the ag-ui-protocol/ag-ui repository added 393 stars in a single day, according to GitHub Trending — a notable spike for a project that has quietly become plumbing.

The pitch is one developers already know from MCP: a front end written against AG-UI can swap the agent runtime behind it without rewriting its streaming layer. Historically each agent framework shipped its own WebSocket protocol and its own React hooks, which made the UI the least portable part of an agent app. If AG-UI keeps spreading, that seam becomes the same kind of commodity the tool seam already is — and the interesting work moves back up into the agent itself.