Skip to documentation
DocsResults and operations

Monitor Experiments and Runs

Follow a comparison plan, inspect one agent Run, and locate the failing layer.

On this page

Experiment view versus Run view

SurfaceUse it for
ExperimentCompare every agent member over one frozen Task selection and see matrix-level progress.
Registry Run DetailRead the authoritative whole-Run aggregate when available, page through rollout rows, inspect the operational journal, and identify an attached immutable report snapshot.
Runtime Run reportRead the member Run's immutable terminal snapshot and complete report statistics.
RolloutRead one Task slot, its Attempt lineage, trace, evidence, and verifier results.

Know whether work is progressing

  1. 1
    Read the rollout states

    At rollout level, Pending, Claimed, and Running are nonterminal execution states. Evaluating is an unresolved Registry UI projection after Runtime execution completes; it does not assert that a Verifier is still running. None of these labels proves the model is generating at that instant.

  2. 2
    Separate the aggregate from the row page

    Run Detail calculates status, progress, evaluation, and failure totals from the authoritative whole-Run summary when Registry can load it. The rollout table is paged independently and clearly labels a partial row view; use Previous and Next or the CLI when you need every individual rollout record. If the aggregate is unavailable, the UI says so instead of deriving whole-Run totals from one page.

  3. 3
    Open the journal

    Claims, results, retries, and the first classified failure show whether Runtime is advancing the Run.

  4. 4
    Treat external trace state independently

    Pending is a normally short-lived indexing state with automatic retry; it does not by itself mean execution is stalled. The trace pane shows the source reason and retry timing. If Pending persists after the Attempt is sealed, treat it as a trace-delivery or indexing diagnostic.

Use the journal to find the failing layer

The Run journal records lifecycle events such as admission, claims, results, failures, retries, cancellation, finalization, and report publication. A failed Run is not automatically an agent-quality failure; start with the first classified failure, then open the Attempt trace for image, Environment, harness, sealing, and verification detail.

Monitor from the CLI

Terminal
fleet-runtime runs status run-... --target fleet-alpha
fleet-runtime runs show run-... --rollouts --target fleet-alpha
fleet-runtime runs show run-... --journal --target fleet-alpha