Arize AI
@arizeai
The AI engineering platform for teams shipping reliable AI agents and LLM applications. Also home to @ArizePhoenix.
Uber spent more than a year studying how to make agent evals useful at production scale. One of the biggest lessons: production should continuously improve your offline evals. That means tracing from the first deployment, turning reviewed failures into new test cases, and- Model pricing pages tell you what tokens cost, but production economics depend on how much work those tokens actually complete. ICYMI from our Arize × @FireworksAI_HQ session: we measured cost per successful task across 2,400 traced runs on Terminal-Bench across K3, GPT-5.5, and
- You can choose a frontier model and still ship an unreliable agent. Production behavior also depends on the context the model receives and the harness that turns its reasoning into action. We teamed up with @AtlanHQ to break down how to evaluate all three layers:
