GenAI-Driven Multiagent Systems for Design to Manufacturing Workflows

Abstract

Recent advances in Generative AI and Large Language Models are creating new opportunities for automating design-to-manufacturing workflows in Computer-Aided Systems. While most prior work focuses on single-agent code generation, multi-agent frameworks that emulate engineering roles with solid models remain limited. This paper presents a lightweight LLM-driven multi-agent prototype for early mechanical design and manufacturability assessment, using Mistral-based agents orchestrated through LangGraph. A designer agent generates OpenSCAD models from natural language, while a manufacturing agent evaluates feasibility and requests revisions. Four use cases, simple geometry creation, infeasible-feature rejection, gear generation with RAG, and iterative co-design of a plate, demonstrate reliable CAD synthesis and meaningful manufacturability feedback. Identified limitations include prompt dependency, reduced fidelity for complex geometries, and scalability constraints. Overall, the study highlights the potential of retrieval-augmented multi-agent workflows for future CAx integration and outlines pathways toward domain-specialized models, richer datasets, and rigorous benchmarking for industry-ready systems.