Walkthrough: handle a failed or blocked agent run

Use this walkthrough when Command Center shows a failed run, a stuck run, or an agent action that was blocked by policy or budget.

Real-life example

An RFQ draft agent was scheduled to prepare a quote packet for Metro Equipment, but Command Center shows the run failed. The operator needs to know whether to retry, cancel, or fix setup before trying again.

Start from Command Center

1. Open Command Center and filter to Failed runs (or Stuck runs).
2. Select the item and read Why this matters.
3. Check Evidence and provenance for the agent, the record it was working on, and the failure reason.
4. Open the record or source from the item to see whether the underlying data is the problem. Do not ask for a retry from habit.

Check recent activity

Open AI OS Activity & Health at /ai-os and look at Recent activity and Loop health for the same loop:

• is this one run failing, or every recent run of the loop?
• are Usage limits close to their ceiling?
• do the Guardrails block the action the run was trying to take?

Advanced detail: the run's event log, checkpoints, policy decisions, budget usage, and the retry and cancel controls are in the internal admin surfaces your Vertiqa team manages. Share the Command Center item with them when a run needs to be retried or cancelled.

Retry, cancel, or fix

Ask for a retry when the failure was temporary: dispatch failed, a provider was unavailable, or a dependency has recovered.

Ask for a cancel when the work is no longer needed, the wrong record was targeted, or repeated retries would create noise. Dismiss the item in Command Center once it's handled, with a reason, or Snooze it while you wait.

Fix setup first when the failure came from missing data, a guardrail, a usage limit, malformed input, or an unavailable connector.

Expected result after action

After a retry, a new run shows in Recent activity. If it succeeds, Command Center stops showing the same failure. If it fails again at the same step, escalate rather than retrying repeatedly.

After a cancel, the run shows as cancelled and the audit trail records who did it.

Failure paths

If Recent activity shows no runs for the loop at all, the problem may be dispatch or scheduling. Check Recurring checks and whether other loops are moving.

If a guardrail blocked the run, loosen it only when the attempted action is safe for that agent and record type.

If a usage limit blocked the run, review whether the output is worth the spend before raising the limit.

Success criteria

The run is no longer stuck without an owner decision. It is either retried, cancelled, or waiting on a specific setup fix, with run history and audit context preserved.

Related

• Agent run recovery
• Policies, budgets, and schedules
• Command Center getting started