Past runs and results

Past Runs separates invocation-level progress from individual run evidence. The summary is suitable for polling; detailed transcripts and frozen definitions are loaded only when you open one run.

The internal conversations created for Agent Testing do not appear in the regular Conversations list or Agent version statistics. Post-conversation Analysis is set to skipped; Agent Testing's own evaluators, parameter matchers, simulator, and failure bucketing provide the result and cost evidence instead.

Read the invocation summary#

An invocation moves from pending or running to a terminal summary after its runs finish. Each run is pending, running, passed, failed, or interrupted. The pass rate uses passed plus failed runs as its denominator; interrupted runs are reported separately and do not silently count as a behavioral failure.

Repeated invocations can include failure groups. Bucketing analyzes terminal failures after execution and stores group labels, run membership, and safe rationales. Raw run results remain the source evidence if bucketing fails or returns an invalid shape.

Page through runs#

The run list is cursor-paginated and can request only rows updated after a timestamp. Summary rows exclude transcripts and frozen definitions to keep polling bounded. The run detail includes result, transcript, events, Tool evidence, criteria, latency, input and output tokens, and cost.

Inspect failure_code before interpreting a failure. For example, quota_exceeded means execution stopped at a model boundary, not that the Agent answered incorrectly. A run may also be interrupted by durable recovery when its previous worker can no longer prove exclusive ownership.

Compare and resubmit#

Resubmitting selected runs creates a new invocation; it never edits the old one. By default, it reuses the prior frozen Agent and Tool snapshot. To compare with current behavior, explicitly choose the current Main Draft or Live target. Up to 5,000 unique run IDs can be resubmitted at once.

Use Durable execution and recovery for interrupted work and bucketing recovery, or Troubleshooting for stable error codes.