I work on Promptfoo (an open-source eval framework). Appreciate the mention here. This post captures a lot of the hard lessons around agent evals. In particular, task ambiguity and brittle graders are things we run into constantly.