Increasingly convinced writing evals is a core competency for working with agents We're used to computers being binary pass/fail on their behaviors but LLMs are more like humans and probabilistic There's a missing tool for making evals understandable for humans