AI agents risk duplicate actions without idempotent tool calls

AI agents performing tasks like refunds can duplicate actions if a network timeout occurs after a service processes a request but before the agent receives confirmation. This problem stems from the execution boundary between the agent and external systems, not the model's reasoning. Retry mechanisms alone are insufficient without idempotency, which ensures repeated requests produce the same effect as a single one. Systems must use unique operation keys to distinguish retries of the same logical action from new requests to prevent unintended duplicates.