As LLMs evolve into multi-step agents, evaluation must move beyond input-output correctness. This session shows how to assess both node-level decisions and full agentic flows, with a live demo of multi-node agents and practical eval techniques.