Follow a task
Runs and Data
Move from live task evidence in Runs to review, grading, and data preparation.
Runs answers what happened during a task. Data helps you understand the evidence, record the behavior you wanted, and prepare improvements. Both read the same retained session ledger under your workspace and project permissions.
Follow the task in Runs#
Select the deployment and find the session ID your application persisted. Open the run to inspect its task, governing revision, ordered requests, review decisions, releases, holds, and application-reported outcomes.
The gateway records supported traffic automatically; you do not need to upload each turn or set up a second webhook. External actions become visible through supported executor authorization and reporting hooks. Activity in Runs is evidence to inspect, not permission for a tool to execute.
Work with the evidence in Data#
Open the source evidence for a decision, then use available review and grading controls. Keep the original session, operation, governing revision, and source record attached to your feedback. Retention and access rules determine what remains available.
| In Runs | Continue in Data |
|---|---|
| The agent proposed deleting an authorization test. | Review the proposal and the task evidence that preceded it. |
| Integrity withheld the proposal or requested a correction. | Compare the decision with the constitution and your expected behavior. |
| The correction preserved the test and earlier implementation work. | Grade both constraint adherence and useful task progress. |
| Similar failures recur across tasks. | Group representative examples and identify missing coverage before generation or adaptation. |
Display updates can lag behind the committed ledger. Reconnecting reads existing evidence; it does not repeat the model operation. An unavailable source should remain unavailable, rather than becoming an assumed successful result.
Choose the next action#
Read a trajectory
Follow a proposal from request through review and release.
Assess behavior
Use custom rubrics and feedback to explain an expected decision.
Fill a coverage gap
Generate variations around sparse, representative examples.
Evaluate an improvement
Understand data preparation, customer versions, activation, and rollback.
For an active hold, inspect the affected operation before changing configuration. For an uncertain external effect, reconcile your executor ledger. The recovery guide explains how to preserve the original work.