The A.I. Daily
Arize demos grading Claude agents with Phoenix
Arize's Laurie Voss demoed Phoenix evals for Claude agents: OpenTelemetry traces log each step, code checks verify deterministic results, and LLM judges grade meaning against rubrics. Testing a Haiku 4.5 to 5.5 upgrade, faithfulness improved over threefold but cost quadrupled; he stressed that "looks good to me" is not a test suite.