Anthropic: Demystifying Evals for AI Agents

4 pointsposted a month ago
by Bayram

1 Comments

dangelosaurus

a month ago

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.