Streaming an agent into FastAPI

A support answer that takes four seconds to appear all at once feels broken. The same answer arriving word by word from the first half-second feels alive, even though it finishes at the same moment. That is why Relay's /chat endpoint streams, and it is why lesson 08.04 taught you to stream a single model call. An agent run is harder, because it is several model calls with tool calls and possibly a handoff in between, and the client wants to know about those too: "looking up your account", "transferring you to billing", not just silence and then text.

Runner.run_streamed gives you all of it as one async stream of events. Some events are raw token deltas from the model. Others mark a tool call starting, a tool result arriving, or the active agent changing. Your job in the endpoint is to pick the events the client cares about, turn each into a line of server-sent events, and yield it.