Inspect Failed Agent Runs in Execution History

A failed run often surfaces first as an unexpected token spike or a stalled schedule. The Agent Command Center surfaces these events in one place so you can inspect failed agent runs without leaving your control plane.

Key takeaways

• Execution history stores every log, intermediate output and cost estimate for later review.
• Error traces link directly to the versioned config object that produced the run.
• Human approval remains required before any retry touches your runtime backend.

Common Reasons for Failed Agent Runs

Most failures trace to one of four categories. Review the list below before opening any execution record.

• Tool permission mismatch in the single config object
• Autonomy level set too high for the scheduled task
• Token usage exceeded the limit defined for that agent
• Model parameter change not yet approved through the approvals inbox
• Execution timeout reached on the agent runtime backend
• Missing intermediate output due to a downstream service error

Failures in these categories often repeat across similar schedules until the underlying config object is adjusted. Teams that track patterns in execution history reduce repeat incidents by documenting each root cause against the active version.

How to Inspect Failed Agent Runs from Execution History

Follow these steps to locate and open a failed run.

1. Open the Agent Command Center dashboard.
2. Filter the history view by status equals failed.
3. Select the run ID that matches the reported schedule.
4. Expand the error trace section to view the stack.
5. Note the config object version active at execution time.
6. Check token usage and cost estimates for that run.

Document each step in your incident log so later audits can reconstruct the exact sequence without re-querying the runtime backend.

Reading Execution Logs and Error Traces

Execution logs arrive in chronological order with line numbers. Error traces include the exact tool call that triggered the failure.

• Timestamp of each log line
• Agent role prompt active during the step
• Tool name and parameters passed
• Returned error code from the runtime backend
• Intermediate output captured before the failure

Compare these fields against the version history of model parameters in config to see whether a recent change introduced the issue. Review Version History of Model Parameters in Config for the exact navigation path. Consistent log formats also align with guidance in the NIST AI Risk Management Framework for traceable decision records.

Using Approval Overrides After a Failure

When a run fails on a sensitive action, the system routes the record to the approvals inbox. You can apply an approval override only after inspecting the full trace.

• Approve the original output with a note
• Edit the output before re-submission
• Reject and require a new config object version
• Route the decision to another team member

Approval overrides remain logged against the run ID for later audit. Audit Agent Decision Paths from Execution History shows how to export these records.

Comparing Successful and Failed Runs

A side-by-side view helps isolate configuration drift.

Use the table above as a template when documenting your own runs. Teams that maintain these comparisons over multiple quarters identify recurring permission or timeout issues faster.

Analyzing Token Usage and Cost Estimates in Failures

Token spikes frequently precede or accompany failures. Review the per-step breakdown to determine whether the agent entered an unexpected loop or called an expensive tool repeatedly.

• Compare observed token counts against the limits set in the active config object.
• Identify steps where intermediate outputs grew larger than expected.
• Note any model parameter version that increased context length.
• Calculate projected cost impact before approving a retry.

This analysis pairs with live visibility features so you can set tighter thresholds in future config object revisions. Track Token Usage Multiple Agents from One Control Plane provides additional guidance on fleet-wide monitoring.

Integrating Checks with Your Runtime Backend

Your agent runtime backend exposes additional fields that the control plane surfaces automatically.

• Network response codes from external services
• Memory usage at each step
• Schedule adherence metrics
• Permission checks performed before tool execution

Secure Agent Runtime Backend with Config Permissions explains how to tighten these checks at the config level. Aligning backend logs with external standards such as those from OWASP improves cross-team incident response.

Best Practices for Reviewing Failures

Apply these practices every time you inspect failed agent runs.

• Always open the approvals inbox before retrying any action that touches production data.
• Record the exact config object version in your incident notes.
• Export the full log bundle before clearing the run from history.
• Compare token usage against the limits defined in Set Tool Call Limits in Your Agent Config Object.
• Reproduce the failure in a non-production schedule first.

These steps keep every change traceable and ensure human approval gates remain intact across runs.

Tracing Changes Across Config Objects

Failures sometimes appear only after a parameter update. Trace the change path with these checkpoints.

• Open the config object diff view.
• Filter by date range of the failed run.
• Identify any autonomy level or tool permission edits.
• Confirm the change passed human review before deployment.

Assign Tasks Across Agents with One Shared Config Object demonstrates how shared objects keep these traces consistent across a fleet.

Reproducing Failures Safely on Your Runtime Backend

Before applying fixes in production, reproduce the failure under controlled conditions. Create a duplicate schedule that points to a test agent runtime backend instance.

• Clone the exact config object version that failed.
• Reduce autonomy level to require explicit approval on every tool call.
• Limit token budget to the observed failure threshold.
• Capture identical execution logs for direct comparison.

This controlled reproduction prevents unintended side effects while you validate the corrected config object.

Conclusion

Once you finish inspecting a failed run, decide on the next concrete action.

• Update the versioned config object and test the change on a limited schedule.
• Add an explicit approval rule for the failing tool call.
• Lower the autonomy level until the next review cycle.
• Export the run record for compliance reporting.

Start by filtering your execution history for the most recent failed run and open its error trace.

FAQ

What fields appear in an execution log?

Execution logs contain timestamps, role prompts, tool calls, returned values and token counts for every step.

How long does the Agent Command Center retain failed runs?

Retention follows the policy set in your single config object; most teams keep records for 90 days.

Can I re-run a failed execution without changing the config object?

Yes, but any sensitive action still routes through the approvals inbox before it reaches your runtime backend.

Which secondary keyword relates directly to retry decisions?

Approval overrides let you adjust the outcome of a failed step while preserving the original error trace.

Where do error traces link back to configuration?

Each trace references the exact version of the config object active at the time of the run.