Evaluating a single LLM call is hard enough. Evaluating an agent that plans, calls tools, revises its approach, and returns an answer three steps later is a different problem entirely. In this session, Chhayank Jain will cover what a practical eval framework looks like for agentic systems in financial services, from trajectory evaluation and LLM-as-judge pipelines to the observability layer you need to catch silent degradation before your risk team catches it first. Key takeaways: - Why pass/fail unit tests break down for multi-step agentic workflows - How to build trajectory evaluation pipelines for non-deterministic systems - Observability patterns that catch silent degradation in production - What a minimum viable eval stack looks like for regulated financial environments