DesAgent: A Multi-Agent Mechanical Design Method Based on Collaborative Large and Small Models

Abstract Mechanical design today faces critical challenges in design efficiency and multidisciplinary optimization, often constrained by high computational costs and fragmented processes. To address these issues, this article proposes DesAgent, a multi-agent collaborative design methodology that integrates the semantic reasoning capabilities of large language models (LLMs) with the numerical prediction accuracy of reduced-order small models (ROSMs). The proposed approach constructs a semantic-numerical synergy loop, enabling a closed-loop, intelligent design process that bridges semantic interpretation and numerical validation. DesAgent features a hierarchical multi-agent system consisting of four specialized agents—requirements analyst, task planner, designer, and feedback evaluator—each responsible for a distinct phase of the design pipeline. The LLMs support natural language parsing and task planning, while the ROSMs ensure real-time simulation-level predictions through neural network-based surrogate models. To validate the proposed methodology, a case study on the structural optimization of a spinning frame wall plate is conducted. Experimental results show that DesAgent reduced material consumption by 21.2% while satisfying multiple constraints related to stress, deformation, and natural frequency avoidance. The entire design optimization process is completed in 232 s, consuming only 12,044 tokens of computational resources. This work presents an efficient, low-cost, and generalizable design framework that demonstrates the feasibility of LLM-augmented collaborative intelligence in complex mechanical design tasks.