An idempotency decision tree for agent steps to avoid duplicate charges

On November 7, 2026, the AI section of DEV Community published a piece titled "Two Writes, One Charge: An Idempotency Decision Tree for Agent Steps". The article opens with a nightly job that walks an invoice queue and calls a payments tool once per invoice. A full run takes eleven minutes. In minute nine the worker is restarted — the reason, the author notes, does not matter: a deploy, an OOM kill, or a dropped connection all produce the same outcome. The retry loop resumes from the last checkpoint and re-executes the step that was already in the air. The visible text of the article cuts off there, but the summary spells out the consequence: two writes, one charge, and one long morning for whoever has to reconcile the account. The author's central argument is that none of this is fixed by a better prompt. What fixes it is deciding, for each step, what "already done" actually means — and encoding that decision by choosing among four mechanisms: an idempotency key, a local deduplication ledger, a compensating action, or doing nothing at all. The choice depends on the step in question rather than on a single global policy. The excerpt available ends mid-sentence at "re-executes", so the full walkthrough of how the decision tree is applied step by step is not visible in the material provided. As of now this appears to be the only report on the topic, with no follow-up coverage, community discussion, or vendor response recorded.