Arize shows how to test an AI agent with traces and judges
In its example, the reports that passed the test went from 5 of 16 to 48 of 48 after a change.
Laurie Voss of Arize shows that teams must record each step of an agent. Teams must score the records with code checks, AI judges and human labels. No other test result is given. For teams that make agents, this gives a number, not a guess.
Sources
Posted
