Applied AI
Agent retries can duplicate a committed customer action
Make side-effecting agent tools safe when a response is lost after the external system commits.
November 10, 2026
An agent may see a network failure even when the remote system has accepted the action. A LangGraph issue reproduces a ToolNode that posts an order, loses the response and retries. The second attempt commits another order. The issue concerns retry behavior, but the underlying ambiguity exists in any at-least-once network workflow.
Distinguish observation from outcome
Log a stable operation ID before the tool call. If the client times out or receives a transport error, query the destination by that ID before sending another write. A missing response is not evidence that the write failed. LangGraph persistence guidance explains checkpointed execution; a checkpoint cannot reverse a side effect already committed outside the graph.
Give read-only tools and mutating tools different retry policies. For a charge, order or CRM update, require an idempotency key accepted by the destination or build a dedupe ledger at the integration boundary. Test the exact failure: server commits, then connection closes before the client sees success.
Define an operator recovery path
When the outcome is uncertain and the destination cannot be queried, stop automatic retry and surface an exception for a person to reconcile. Preserve the tool input, operation ID, upstream response if any, and graph checkpoint. The customer-facing outcome is one correct action, not an agent run marked successful.
Adovo’s view
An intelligent agent is still accountable for the writes it makes. Retry is a transport convenience; correctness is a business contract. Give every consequential action an identity that survives the network.
