Streaming and structured responses

An agent run takes seconds, sometimes tens of them, and two readers are waiting on it. One is a person watching the Relay chat window who needs to see something within the first half second. The other is your code, which needs the result in a shape it can act on: a category, a flag, a ticket id. Streaming serves the first reader. Structured responses serve the second. A good /chat endpoint does both, and keeps them separate.

Both come from the same runtime. Every LangGraph graph, including the one create_agent builds, has stream and astream methods that yield events while the run is in progress, and create_agent takes a response_format that makes the final state carry a validated object.