Recover and export executions
Choose the correct recovery action for an execution, then verify the result or export its trace as JSON.
Use an execution's status and your intended outcome to choose between cancellation, safe-step retry, replay, approval-driven resume, and JSON export. This guide helps you avoid repeating work when recovery from the failed step is available.
Before you begin
You must be signed in and able to access the execution. Start from the execution detail page for one of your runs; Run and find executions explains how to locate it.
Choose an action
The actions have different effects on execution history:
| Action | Use it when | Result |
|---|---|---|
| Cancel | The status is RUNNING or PAUSED_FOR_APPROVAL and the current run must stop | The current execution is cancelled |
| Retry safe step | The status is FAILED, CANCELLED, or STEP_LIMIT_EXCEEDED and completed safe steps should not run again | Recovery starts from the failed point in the same execution context |
| Replay / rerun | You want to run the original skill version and input again | Agent Studio creates a new execution linked to the original |
| Resume | A human decision gate has been approved | The paused execution continues from its restored state |
| Export JSON | You need the complete report for analysis or retention | The browser downloads a JSON file without changing execution state |
Retry continues from the failure boundary, whereas replay preserves the historical run and creates another linked run. Resume is not a general-purpose retry: it requires an approval in APPROVED state and uses the approval's idempotency key to prevent duplicate continuation.
Recover or export a run
Open the execution detail page and review its status, input, completed steps, and failure information. The header shows the run status and exposes only the recovery controls that apply to that state.
If the cause is unclear, inspect the trace before changing state. See Inspect an execution trace.
For a run with status RUNNING or PAUSED_FOR_APPROVAL, select [ CANCEL RUN ] only when no further workflow work should continue.
After the request succeeds, Agent Studio displays Execution cancelled and refreshes the execution data. Cancellation applies to the current execution; use replay later if you need a new run with the original input.
For a run with status FAILED, CANCELLED, or STEP_LIMIT_EXCEEDED, select [ RETRY SAFE STEP ] when completed safe steps should remain completed.
Agent Studio rejects this operation if the execution is already running. A successful request displays Safe Step Recovery Started and resumes directly from the failed step without repeating completed safe steps.
Select [ REPLAY / RERUN ] when you need a complete new execution using the historical run's skill version and input.
Agent Studio creates a new linked execution, leaves the original record unchanged, and opens the new execution detail page. The new page identifies the source with a REPLAYED FROM label.
Open Human Review Queue, inspect the pending request, and select APPROVE & RESUME only after the requested action is acceptable.
Approval records the decision first. Agent Studio then supplies the matching approval ID and idempotency key to resume the paused execution. For the complete decision workflow, see Review and cancel approval requests.
Select [ EXPORT JSON ] to download the execution report. The file is named execution-<execution-id>.json and contains the trace data, structured logs, timeline events, and approval records available for that execution.
Agent Studio displays Export downloaded after the browser starts the download. Export is read-only and does not cancel, retry, replay, or resume the run.
Verify the result
Confirm the outcome that matches your action:
- cancellation changes the current run from an active state;
- retry reports that safe-step recovery started;
- replay opens a different execution with a REPLAYED FROM label;
- approval reports Action approved and Execution is resuming;
- export creates the expected
.jsondownload.
Troubleshooting
Next steps
Use Inspect an execution trace to analyze node steps and logs, or Use the Observability dashboard to compare reliability and usage across runs.