AG-UI: The Missing Protocol Between Agents and Users

The agentic protocol stack has grown up in layers, but one layer has been conspicuously underspecified: how agents actually communicate with the humans they're supposed to be helping. MCP handles agent-to-tool interaction. A2A handles agent-to-agent coordination. But the agent-to-user channel — the moment where an agent surfaces a result, asks a clarifying question, or receives a mid-task correction — has been largely ad-hoc, varying by framework and implementation. AG-UI, the Agent-User Interaction Protocol, proposes to fix that.

Introduced in a new slide deck by Mete Atamel from Google DevRel, AG-UI defines 16 standardized event types for bidirectional, event-driven communication between agents and users. Agents emit structured events as they execute — progress updates, intermediate results, clarification requests — and users can inject inputs back into the running loop without interrupting or restarting it. The protocol is deliberately transport-agnostic, supporting SSE, WebSockets, and webhooks, and the deck includes a live demo of AG-UI running with Google ADK at dojo.ag-ui.com.

As streaming, human-in-the-loop workflows, and collaborative agent tasks become table stakes rather than differentiators, having a formal spec for this interaction layer matters. Right now, every framework reinvents its own user-communication model; AG-UI is a credible attempt to give that problem a shared vocabulary and a stable interface — enabling consistent UX patterns regardless of which underlying agent framework is doing the work.

Read the full article at Speaker Deck (Mete Atamel, Google) →