Skip to main content
Results from every evaluation, hosted or from your own worker, land in the same places.

​
Compare scores over time

Go to Observe → evaluations.
  • Recent runs lists each evaluation as it lands: whether it came from a hosted (managed) or your own (customer) evaluator, the agent and session, the evaluation and its version, its status, and its score or metrics.
  • Score over time plots what you ask for. Select add series and choose an agent, an environment, an evaluation, and a statistic: avg, min, max, p50, p75, p90, p95, p99, stddev, or mode. Each series is one line; give it its own curve to draw it on a separate chart. The evaluations page: recent runs tagged customer, a score over time chart with reference lines at 0.5 and 0.8, and one series averaging finished_clean across all agents and environments.
One time range and one bin size apply to every series. A fine bin finds an incident; a coarse one shows a trend, and can hide the spikes you are looking for. A bucket where nothing was scored is a gap in the line, never a zero, and reference lines mark 0.5 and 0.8.Every part of the view lives in the URL: share copies it, and whoever opens it sees exactly the comparison you built.
Plot avg and p90 for the same evaluation to see whether a good average is hiding a bad tail, or the same evaluation for two agents, or for production and staging, to compare them on one axis. Costs, latencies, and token counts, which carry units, chart under Observe → metrics, one chart per unit.

​
See why a session scored low

Open a session from Observe → sessions; the grid carries each session’s scores and filters by score range. The session’s right rail leads with the evaluation summary, then a bar per score with the evaluator’s reasoning under it. A session detail view showing evaluation scores and reasoning beside the complete trace.

​
Ask the assistant

Ask about evaluation data in plain English: “tell me about some of the recent evaluations”, or which agents’ scores are slipping. The assistant reads and analyzes the results and answers with tables you can follow up on, and a question worth keeping can become a query or a dashboard. The evaluations page beside the assistant, which answers "tell me about some of the recent evaluations" with a summary of totals, statuses, and scores.

​
Watch and act

  • Dashboards, under Analyze → dashboards, trend the scores you feature, per agent and environment, for the whole organization. A quality dashboard showing average evaluation scores and trends over time.
  • Alerts notify you when a score crosses a threshold. See alerts.
  • When a score declines across many sessions, run an audit to find out why; when the cause is a repeatable action, write a policy.