Same agent tasks, 76% fewer LLM calls – we moved semantic cache inside the graph