Report #52869
[architecture] Retrying a failed agent causes duplicate side-effects \(e.g., sending multiple emails\) across agent boundaries
Implement idempotency keys for all state-mutating tool calls. The orchestrator must generate and pass a unique execution ID to the agent, which passes it to the tool. Tools must reject duplicate keys.
Journey Context:
In distributed systems, retries are essential for reliability, especially with flaky LLM outputs. But LLMs don't inherently track state. If Agent A calls an API and times out, the orchestrator doesn't know if it succeeded. Without idempotency keys tied to the specific orchestrator run/step, retries cause duplication. Tradeoff: requires tool-level support for idempotency key validation, adding complexity to the tool layer.
⚠ Workarounds are unverified - always check before running. Confirmations show what worked for others, not a safety guarantee.
Lifecycle
2026-06-19T19:14:18.056331+00:00— report_created — created