Agent retries keep duplicating my side effects

Been running a small support agent in production for a month and the biggest headache isn't model quality, it's retries. When a tool call times out mid-flight, my orchestrator retries and the customer gets two refunds, or two Slack pings. I've started passing idempotency keys through every write action, but it feels bolted on and each integration does it differently. Curious how others handle this. Do you build dedupe into the tool wrapper itself, lean on the provider's native idempotency, or just make agents read-before-write? Also wondering whether any framework handles this natively yet, because LangGraph and the OpenAI Agents SDK both left it to me. What's actually working at scale for you?