Review costs, token usage, and operational success and error trends across the runs available in Studio. Use model, agent, and workflow breakdowns to prioritize what your team should investigate and refine.
Operational success is not a measure of output accuracy. Assess the result against your own requirements.
Conceptual overview. Not a live product screen.
Working with Evaluation
Measure operations. Apply your judgement.
Review operational performance
See costs, input and output token usage, and success and error rates for the available run data. These measures describe execution, not the correctness of the result.
Look at the breakdowns
Review performance by model, agent, or task. Average turns and duration help you identify work that deserves a closer look.
Connect metrics to the trace
Use Observability to investigate the recorded work behind a pattern. Decide whether the next change belongs in instructions, tools, or model configuration.
Improve deliberately
Make a focused change, run the work again, and inspect the result against your requirements. Your team remains responsible for assessing output quality.