Skip to main content
Guardrails describe topics or behaviors that need a defined response. Start by writing what the agent may help with, when it should ask for clarification, and when it should stop or direct the user elsewhere.
The dashboard’s Guardrails page is a preview: changes there are not published. The API includes guardrail authoring fields, but saved settings do not establish live moderation, prompt-injection protection, or automated intervention.

​
Write the policy first

For each boundary, keep a realistic example, the expected response, and the service responsible for enforcing any access restriction. For example, a support agent can explain the account-recovery process while the account service verifies identity before changing a login method. Do not rely on a prompt or guardrail setting to grant or deny access to private data. Tool authorization must be enforced by the service that owns the action.

​
Review the authored settings

Guardrail configuration lives under platform_settings.guardrails. The current version discriminator is version: "1". Content categories include sexual content, violence, harassment, self-harm, profanity, religion or politics, and medical or legal information. Category entries contain is_enabled and threshold. A stored threshold is a configuration value, not a measured accuracy or protection guarantee.

​
Prepare a review

1

Choose relevant boundaries

Select categories that matter for the task. Include ordinary requests that must remain answerable, so the review covers both inappropriate responses and unnecessary refusals.
2

Save API-authored configuration separately

If preparing guardrail fields through the API, preserve the complete agent snapshot and use a server draft. Read the draft back. Do not treat changes in the preview dashboard as a saved policy.
3

Record the expected response

For each example, specify acceptable wording, any handoff, and whether an external action must be denied. Keep execution verification pending until the applicable runtime path is available.

​
Frequently asked questions

No. Saving a field and verifying runtime enforcement are separate checks. Do not describe the agent as protected against a class of attacks based only on its configuration.
The Guardrails page is labelled as an unpublished preview. Use the API draft workflow for configuration authoring, and inspect the returned document.