"Ambiguity and mistakes can be carried forward through multistep agent workflows, allowing incorrect assumptions to influence later decisions."
Zhu and colleagues examine failures in multistep language-model agents, including errors that propagate across successive actions. Their work supports the concern that agent workflows can carry forward earlier misunderstandings, making error detection and feedback important parts of reliable autonomous execution.
View Source