What kind of Agent validation are you doing today?

Everything started with model Evals and benchmarks (which model is better?), then evolved to prompt management and from there to analyzing traces. What do people do today, and how are they sourcing test datasets?