Configure Execution Guardrails

Execution Guardrails detect when a specialized agent run behaves unlike a normal run for your agent. Use Execution Guardrails to scale specialized agents safely across your organization without manually reviewing every execution.

Configure Execution Guardrails per agent on the Execution Guardrails sub-tab of the Quality tab.

Execution Guardrails outcomes are independent of Output Evaluation results. Execution Guardrails sometimes flag a run that Output Evaluation marks as Passed, and vice versa. See Configure Output Evaluation for details on quality scoring.

Open the Execution Guardrails sub-tab to configure guardrail settings or view results for a specialized agent.

Execution Guardrails progress through three modes as Opal learns your agent's behavior. The agent's current mode displays the Current label next to the mode name on the mode description card. Use the navigation arrows to view the description of any mode.

Only successful runs advance an agent through the modes. Failed runs and rejected runs do not progress the agent toward the next mode.

Learning mode covers the first 20 successful runs of an agent version. Opal observes runs to build a behavior baseline. Opal does not block any runs during Learning mode.

Watching mode covers the next 30 successful runs after Learning mode ends. Opal flags unusual runs as warnings so you can review them. The baseline continues to sharpen during Watching mode, but Opal does not block any runs.

Enforcing mode activates after Learning and Watching modes complete. Opal stops runs that look clearly unsafe compared to learned behavior and alerts you when this happens.

Opal categorizes each run by outcome and displays the breakdown on the Summary card. The bar on the card visualizes the proportion of each outcome using the same color coding as the legend.

The Summary card displays the following outcome categories:

A Warning or Failed outcome means the agent completed the execution but Opal detected unusual output patterns. Investigate the agent's inputs, configuration, and dependencies to find what changed.

Warning outcomes indicate a moderate deviation. Common causes include the following:

Failed outcomes indicate a significant deviation that typically requires investigation. Common causes include the following:

Opal sends a notification when a run produces a Warning or Failed guardrail status. See Notifications overview.

The Guardrail settings section tunes how Opal flags and stops runs. The section contains two controls: the Sensitivity slider and the Automatically stop runs that breach guardrails toggle.

The Sensitivity slider controls how aggressively Opal flags or stops runs that fall outside the established baseline. The slider ranges from 1 (Lenient) to 5 (Aggressive). The default is 1.

A lower setting flags fewer runs and is less likely to catch subtle anomalies. A higher setting catches more anomalies but also flags some runs that are slightly unusual but ultimately safe.

When you change the Sensitivity setting, Opal re-evaluates past runs against the updated value. Outcome counts on the Summary card update to reflect the re-evaluation. The updated setting also applies to future runs.

Toggle Automatically stop runs that breach guardrails on to have Opal stop runs that breach the baseline in Enforcing mode. The toggle has no effect during Learning or Watching mode. Opal never blocks runs in those modes regardless of the toggle setting.

Toggle it off to have Opal alert you when a breach happens without stopping the run.

The toggle is on by default.

Excluding individual runs from the baseline lets you keep the training data clean. This is useful when an outlier or test run skews what Opal considers normal.

Reject runs in the following situations:

Rejection works for any run regardless of its Guardrail Status.

The Guardrail Status field displays a Rejected label next to the original status, and the button changes to Approve.

After approval, the Rejected label no longer displays and the button changes back to Reject.

Inheriting the baseline prevents your agent from restarting Learning mode each time you publish a version. When you create a version of an agent, Execution Guardrails carry forward what they learned from the previous version. Learning mode does not restart.

A banner on the Summary card identifies which version the inherited baseline came from. The current version continues to refine the baseline as additional runs accumulate.