Getting Started With (Offline) Agent Evals

We built an eval for our agent's tool selection, learned it was measuring the wrong thing, and fixed it. The three-level framework, the LangSmith code to reproduce each step, and what it cost.