Where to start with agent evals, offline vs online evals, evaluator types, LLM judges, benchmarks with Harbor, multi-turn evals, and keeping costs down.