Agentic Large Language Models for Conceptual Systems Engineering and Design

This study investigates whether large language model (LLM)–powered agents can effectively manage long-context reasoning, ensure task continuity, and perform iterative refinements in early-stage engineering design. Focusing on a Solar-Powered Water Filtration System, we first engaged OpenAI’s o1 model through a human-driven conversation to establish a high-quality baseline for functional decomposition, subsystem mapping, and numerical model scripts. We then developed a multi-agent framework, which leverages a graph-based approach to capture and refine the system’s functions, subsystems, and numerical modeling elements. An ablation study compared this multi-agent setup to a simpler two-agent system. Both agentic frameworks are powered by the LLama 3.3 70B model. Neither approach succeeded in fully recovering the required physics-based numerical models. However, the multi-agent system demonstrated deeper coverage by detailing subfunctions, requirements, and constraints. Despite these strengths, the LLM encountered a failure mode during extended context processing, preventing successful completion of the design workflow. These findings underscore the challenges of handling large, evolving design contexts within current LLM architectures. To address these limitations, improved inter-agent communication strategies and advanced reasoning LLMs—such as OpenAI’s o1 or DeepSeek’s R1—may help enhance design coherency. Further research on robust context management and structured messaging is needed to achieve an end-to-end automated engineering design capability.