Why AI agents need idempotency keys to avoid duplicate writes

An agent that issues a refund and receives a timeout cannot tell whether the payment went through, so a retry may trigger a second refund. The problem stems from infrastructure rather than model reasoning, and it is amplified because agents retry from several layers at once, including the model itself, frameworks, SDKs, and message queues. The established remedy is idempotency: each intended action gets a stable key that is sent with every attempt, and the receiving system returns the recorded outcome rather than executing again. The article notes that keys must be derived from the business identity and that provider-side key retention is limited, so workflows that may retry over longer periods need their own records.