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.

Prerequisites

• A specialized agent exists in your Opal instance. See Create a specialized agent.
• You have permission to edit the agent.

Access the Execution Guardrails sub-tab

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

1. Go to home.optimizely.com.
2. Select your organization.

3. Click Opal.

   

4. Click Agents.

   

5. Click the Your Agents tab.

   

6. Select a specialized agent.
7. Click Quality.

8. Click Execution Guardrails.

   

Guardrail modes

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

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

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

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.

Execution outcomes

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:

• Learning – Executions that occurred during Learning mode and contribute to the baseline.
• Passed – Executions that looked normal compared to the baseline.
• Warning – Executions that deviated moderately from the baseline (50–70%). Opal flags them for review but lets them complete.
• Failed – Executions that deviated significantly from the baseline (over 70%). These show notable differences from normal behavior.
• Terminated – Executions Opal stopped due to extreme deviations from normal behavior. Only happens in Enforcing mode when you enable auto-stop.
• Rescued – Executions Execution Advisor recovered to completion after guardrails would have stopped them.
• Unapproved – Executions Opal excluded from the baseline, either due to failures or manual rejection.

Configure guardrail settings

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.

Sensitivity

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.

Automatically stop runs that breach guardrails

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.

Manage the guardrail baseline

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:

• A run is an outlier that does not represent typical agent behavior.
• A test run produced results that you do not want in the baseline.
• A run was anomalous for reasons unrelated to the agent's normal use.

Rejection works for any run regardless of its Guardrail Status.

Reject a run

1. Open the specialized agent.
2. Click Logs.
3. Click the run you want to reject.

4. Click Reject on the Guardrail Status field in the execution details panel.

   

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

Approve a rejected run

1. Open the specialized agent.
2. Click Logs.
3. Click the rejected run.
4. Click Approve on the Guardrail Status field in the execution details panel.

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

Baseline inheritance

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.

Related articles

• Quality tab overview
• Configure Output Evaluation
• Execution Advisor