Using AG-UI with Foundry Hosted Agents #4720

Hi Viktor Shamal Andersen (@shamal-kp), I ran into the same wall and ended up building a working example, so sharing what worked in case it saves you some time: https://github.com/Santiagosala2/foundry-agui-example

You're right that AgentFrameworkAgent doesn't fit into the wrapper from azure.ai.agentserver.core - that path assumes the responses protocol, which is OpenAI-compatible and platform-managed, so you never get control of the event stream. AG-UI needs its own SSE format, so that's a dead end.

But you don't need the proxy app either. Hosted agents also support the invocations protocol, which hands you the raw HTTP request and lets you write the SSE stream yourself — the docs even list AG-UI explicitly as an invocations use case (protocol table here). So the whole thing runs inside the hosted container, and you keep the observability/session storage/scaling benefits.

The core of it is just an invoke handler that pipes AgentFrameworkAgent's events through the AG-UI encoder:

app = InvocationAgentServerHost()

@app.invoke_handler
async def handle_invoke(request: Request) -> StreamingResponse:
body = AGUIRequest.model_validate(await request.json())

async def event_stream():
    encoder = EventEncoder()
    async for event in protocol_runner.run(body.model_dump(exclude_none=True)):
        yield encoder.encode(event)

return StreamingResponse(event_stream(), media_type="text/event-stream")

where protocol_runner is your AgentFrameworkAgent(agent=..., state_schema=..., predict_state_config=...) — same as in the recipe agent example. In azure.yaml you declare protocols: invocations instead of responses.

On the CopilotKit side, the runtime just points an HttpAgent at the agent's invocations endpoint. Two gotchas once it's deployed: you need a bearer token for the https://ai.azure.com/.default scope (the caller needs the Foundry User role), and a Foundry-Features: HostedAgents=V1Preview header.

My repo has the full thing end to end — a travel planner with shared state (predict_state_config streaming into the UI), the azd deployment setup, and the Next.js/CopilotKit side. Happy to answer questions if you hit anything.