How would you design a fallback mechanism for an AI agent when an external tool or API dependency fails in a production environment like Clear?
