Expected output
TOOL_CALL events arrive before any MESSAGE_DELTA tokens — the model decides to search first, then streams its answer. Both the tool result and the streamed text are LLM-driven and will differ per run.
See also
Multi-turn session
Reuse the same session for follow-ups — the agent remembers earlier turns.
Session send_many
Batch several messages into a single turn before streaming.
Agent reference
Full method surface.
