Definition: A mechanism that limits and controls an Agent’s inputs, outputs, or actions to prevent unsafe or undesirable behavior.
Simply put: A Guardrail tells the Agent: “You can go this far, but no further.”
Examples:
- Preventing execution of
rm -rf. - Preventing the transmission of confidential information.
- Rejecting responses that violate security rules.