防护机制

Guardrails are built-in safety and quality checks that help ensure agents behave reliably, securely, and in alignment with organizational policies. They monitor and control what data agents process and how they respond to user inputs, reducing the risk of exposing sensitive information or executing unintended actions.

In Agents, guardrails can be applied at different levels of execution:

• Agent level: Evaluates the agent’s own prompts or system instructions before execution.

• LLM level: Monitors the requests and responses exchanged between the agent and the large language model.

• Tool level: Validates the inputs and outputs of tools used by the agent during task execution.

By combining these scopes, you can design safer agents that automatically detect and block problematic behavior, maintain compliance, and improve overall trust in automated decision-making.

Go to Out-of-the-box guardrails to quickly set up guardrails at the agent or LLM level.

Go to Custom guardrails to learn how to configure guardrails for your agent’s tools.

Go to Monitoring guardrails to learn how to verify detections and review guardrail activity.

Best practices for guardrails

Follow these recommendations to get the most out of your guardrails:

• Use Block actions in production. For agents that handle sensitive or business-critical data, configure the guardrail action type to block to prevent unintended data exposure or unsafe execution.
• Test thresholds during evaluation. Run evaluation sessions to fine-tune detection thresholds and achieve the right balance between accuracy and false positives.
• Review logs regularly. Monitor guardrail logs and traces periodically to confirm that detections are triggered as expected and adjust configurations if necessary.