AI that can act (call tools, run code, hit APIs) is powerful—and risky. Vibe-checking outputs doesn’t scale when agents plan, execute, and adapt on their own. In this talk, we replace guesswork with evaluations: the crash-tests for agent behavior. You’ll learn a practical toolkit of code-based checks (rules and invariants), human reviews (gold-standard sampling), and model-graded evals (scalable judges with guardrails). We’ll run a live demo: define failure modes, write semantic unit tests, iterate prompts/policies, and wire the evals into CI so regressions get caught before users do. You’ll leave with a small template you can drop into your agent stacks to measure reliability, reduce cost loops, and keep actions safe—moving from “it seems fine” to “we have proof.”