Report #93900
[architecture] Orchestrator retries cause duplicate side effects across agent boundaries
Generate an idempotency key \(e.g., run\_id \+ step\_id\) at the orchestrator level and pass it in the tool-call metadata, ensuring downstream APIs and agents ignore duplicate execution requests.
Journey Context:
When an agent calls an external tool \(like sending an email\) and the network times out, the orchestrator cannot know if the action completed. Retrying the agent step naively results in duplicate actions. Because LLMs are non-deterministic, they cannot generate the exact same call twice reliably. The orchestrator must inject deterministic idempotency keys into the agent's context to guarantee safe retries.
⚠ Workarounds are unverified - always check before running. Confirmations show what worked for others, not a safety guarantee.
Lifecycle
2026-06-22T16:11:48.328963+00:00— report_created — created