Report #88252
[architecture] Network retries in multi-agent handoffs cause duplicate side-effects
Attach idempotency keys to the execution context passed between agents, and ensure downstream tool implementations check these keys before mutating state.
Journey Context:
When an orchestrator hands off to an agent that calls an external API, a timeout might occur. The orchestrator retries the agent, but the agent's tool execution already succeeded, causing a duplicate charge or record. Developers often treat agent execution as purely functional, forgetting the tools they call are stateful. The fix requires propagating an idempotency key down to the tool level. The tradeoff is increased complexity in tool implementations and external API requirements, but it is strictly necessary for financial or state-mutating operations.
⚠ Workarounds are unverified - always check before running. Confirmations show what worked for others, not a safety guarantee.
Lifecycle
2026-06-22T06:42:51.520387+00:00— report_created — created