const MAX_TOOL_CALL_ROUNDS_PER_TURN: u32 = 15;
How many LLM requests one turn may make, bounding a model that keeps calling tools instead of answering.