high75%
Agent factory `_execute_model_async` uses non-streaming `ainvoke`, preventing `on_llm_new_token` callbacks
The agent factory's async model execution path uses `await model_.ainvoke(messages)`, which returns a complete response without emitting token-level callbacks. LangGraph's `use_astream=True` only controls graph-level streaming, so model token streaming via `on_llm_new_token` is never triggered. This breaks real-time token streaming for agents.