Design Patterns for Compound Multi-Agent LLM Systems in Engineering: An Industry Case Study

Le résumé fourni par la source

Large Language Models (LLMs) are increasingly used in engineering, predominantly as standalone models for text-heavy tasks such as requirements analysis and general information retrieval. However, their potential extends further: when integrated with external tools and architected as Multi-Agent Systems (MAS), resulting compound LLM systems can act as intelligent interfaces between users and domain-specific tools, enabling the automation of more complex workflows with various forms of engineering data beyond text. Currently, practical guidance on how to structure and build such systems that integrate effectively with domain-specific workflows is lacking. To address this gap, we present an industry case study of a multi-agent LLM-based assistant tailored for simulation-driven product development workflows. From this case as well as literature insights, we extract design patterns for compound LLM systems. Our approach integrates specialized LLM ReAct tool agents, coordinated by strategic reasoning agents to support simulation analysis tasks using the Ansys Data Processing Framework while incorporating internal material databases and embedding-based component name matching. Our work builds on a theoretical framework for developing LLM-based MAS. We apply its six-step methodology to this industrial use case. It is then evaluated using representative engineering queries derived from workshops with industry practitioners to assess initial performance and interpretability. The outcome is a set of theoretically and empirically grounded design patterns for building compound LLM systems in engineering contexts. These include general best practices as well as hands-on learnings regarding model selection, prompt-engineering, tool-integration strategies, agent coordination and evaluation methods. These findings aim to provide actionable guidance for researchers and practitioners developing compound LLM systems in engineering domains enabling the AI-based automation of more complex tasks.

Ce résumé expose les affirmations des auteurs. BNTIC ne l’interprète pas comme une validation indépendante des résultats.