Skip to main content
Open Analytics in the sidebar. Choose a workflow, scenario, agent and model, product-choice context, and date range. From (UTC) includes that date; Before (UTC) excludes it. The same context follows every result you inspect. Agent and model labels use their catalogue display names. Workflow steps use the frozen definition from the runs, so editing a workflow does not rename an earlier result.

​
Read a result

Results show Scenario-weighted, Pooled, and contributing run counts. Select Inspect runs to see the records behind a result. Open the information controls beside a result for help reading its labels. Completed after reaching the workflow counts completed jobs among runs that reached the job. It is separate from whether the agent chose your product. A normally finished session does not by itself prove a completed job. No measurable results means no eligible runs contributed. It is different from a measured zero. Not enough runs yet marks a result whose uncertainty is still wide; its headline shows run counts. Counts left out of each measure, including failed judgments, stay visible beside the result.

​
Choose a view

  • Overview shows reach beside completion, compares workflows and shows completion over time. Missing results leave a gap in the graph.
  • Choice & discovery shows which products agents discovered, mentioned, recommended and chose, including alternatives and no choice. Reasons against opens the recorded quotes behind objections to a product.
  • Completion & blockers shows workflow progress, surfaces used, completion by agent and model, and groups of observed failures. Product problems appear separately from other recorded failures. A recorded failure identifies what happened; it does not by itself establish why it happened.
  • Steps & cost compares completed jobs against their verified paths. Input, cache read, cache write, output and reasoning remain separate. Usage per completed job includes the attempts that reached the job but did not complete. Not fully priced means some usage has no price; it is not zero cost.

​
Inspect the evidence

Select Inspect runs beside a result. The side inspector shows the counted runs, the runs showing the result, and the counts left out of that measure. Runs awaiting judgment, failed judgments and answers from earlier questions stay separate from counted results. Interrupted runs count only when you select Include interrupted runs. Runs that could not start never enter a rate. Open a run inside the inspector, then select Full run record to read its execution. Quotes and failure examples can open the relevant recorded step. Back in inspector returns to the preceding evidence without losing your analysis context. Results from older setups stay separate from the current result. In Older setup series, inspect individual runs and their frozen context; no combined rate is shown across different older setups. Changing the workflow definition or the model served can start a new setup group; an agent release alone does not. Filter by topic to see results for that need. Earlier runs retain the topic they had when launched, even after a scenario is moved.