The increasing complexity of market environments, characterized by reduced batch sizes and a growing diversity of product variants due to mass personalization, demands greater adaptability and flexibility from production systems. To effectively manage this complexity, material flow simulation can be utilized. However, since simulation software is an expert tool, the entry barrier for inexperienced production planners remains high. This barrier can be lowered by developing virtual assistants that minimize manual effort and support informed decision-making in dynamic production contexts. Recent advances in large language models offer considerable potential for automating tasks, particularly in production system planning and control. However, due to the complexity of the underlaying task, large language model-based single-agent systems often struggle to solve complex, multi-step problems reliably. To address this limitation, multi-agent systems have been proven to be appropriate. Therefore, new system architectures are required for large language model-based multi-agent systems. Despite the relevance of this topic, there is currently a limited amount of research on the design and implementation of such architectures within the context of simulation studies. This approach introduces a multi-agent system architecture that enables the supported execution of material flow simulation studies. The behavior of the multi-agent system is analyzed to understand the potential and challenges of how such systems can serve as virtual assistants for production planners conducting simulation studies. A key application is the integration of these large language model-based agents into digital twins of production systems, thereby enabling time savings through supported model adaptation, lowering entry barriers for inexperienced users, enhancing decision support, and reducing error susceptibility through structured procedures. The findings offer insights into architectural design, tool use, and memory mechanisms that support simulation studies and mitigate hallucinations in large language models.
Korth et al. (Thu,) studied this question.