Evaluating whether an AI agent completes a task is hard. Evaluating whether it does so safely is harder — and most teams have no systematic way to catch failures before users do. What breaks when you try to measure safe behavior across multi-step, non-deterministic agent workflows? Why aren’t classic eval pipelines built for this? And what practical evaluation strategies actually work? This talk tackles all three, offering concrete patterns — from trajectory-level assertions to adversarial scenario generation to safety-aware scoring rubrics — for teams shipping agents today. Drawing on experience building LLM evaluation frameworks and production safety systems at scale, we go beyond surface-level pass/fail metrics to show how you can build real, repeatable confidence that your agent is behaving as intended.