AWS Lambda Durable Functions and the Replay Bug in Human-in-the-Loop Agents

A recent article explores the risks of using AWS Lambda Durable Functions for human-in-the-loop workflows. The author highlights a critical 'replay' mechanism where the Lambda handler restarts from the beginning upon resuming after a pause. If non-deterministic operations—such as AI agent calls—are placed outside of a 'context.step', they may re-execute during the replay, leading to inconsistent outcomes where the system executes an action different from the one approved by the human. The author demonstrates this 'replay bug' and provides a structural fix: wrapping all non-deterministic or side-effecting code within a durable step. The piece also covers common pitfalls, including input type changes and raw byte handling in callbacks. By comparing Durable Functions to AWS Step Functions, the author emphasizes that while Durable Functions simplify code, they require developers to maintain strict discipline regarding replay safety to prevent silent, dangerous logic errors.