Skip to content

Simulations

The Simulations surface shows multi-turn agent simulation runs created by simulate() — a list of runs with pass-rate, and a per-run detail view with each scenario's transcript and per-criterion verdicts.

Find it in the sidebar under // EVALUATION → Simulations. Runs appear here automatically whenever you call simulate() (persistence is on by default).

Simulations list

List page (/simulations)

Every simulate() run, newest first, with:

  • the run name (or id) and the agent under test,
  • the scenario count,
  • a pass-rate bar (green ≥ 90%, amber ≥ 70%, red otherwise),
  • when it started.

Click a row to open the detail.

Detail page (/simulations/{run_id})

A header strip with pass rate, passed / failed counts, and start time, then one card per scenario. Each card shows PASS / FAIL at a glance; expand it to see:

  • per-criterion chips — green when the desired state holds (success criterion met, or failure criterion absent), red otherwise; hover for the judge's reason;
  • the transcript as chat bubbles (user on the left, assistant on the right), with a "view trace" link on each assistant turn that deep-links to the full trace at /traces/{trace_id}.

Simulation detail with transcript and per-criterion chips

Because each agent turn nests under the simulation's trace, you can jump from a failing scenario straight into the trace — and from there into Agent Replay — to debug exactly what the agent did.