Because LLM calls carry uncertainty and the possibility of failure, an Agent scheduler must have built-in retry mechanisms and idempotency guarantees; otherwise, a single network blip could cause the Agent to execute repeatedly and produce irreversible side effects.