Evals

Anyone can build a chatbot. The harder question is whether it’s any good — so the agent has a test suite, and it runs against production, not a mock. Groundedness, tool routing, refusals, and prompt injection.

The last recorded run loads by default. Running it live costs about 20¢ of real API spend, so it’s rate limited. Expand any case to see the exact question, the full answer, and what each grader concluded.