The automation of structural analysis workflows has long been constrained by the need for expert intervention across multiple software tools and iterative decision-making processes. While recent studies have demonstrated the potential of large language models (LLMs) for individual engineering tasks[1–6], end-to-end automation of CAD generation, finite element analysis (FEA), and design optimization remains a challenging problem in computational mechanics. This work presents an autonomous multi-agent LLM framework that enables structural design and analysis workflows directly from natural language specifications. A planner agent coordinates specialized agents responsible for parameter extraction, CAD generation, mesh generation, structural analysis, and design optimization. Design intent expressed in natural language is translated into structured geometric and material parameters, which are used to generate validated CAD models using FreeCAD. Volume meshes are generated using Gmsh, converted into MAPDL-compatible formats, and solved using PyMAPDL with minimal human intervention. A key feature of the framework is its adaptive workflow management based on the internal state of the analysis process. Information obtained during geometry generation, meshing, and solver execution is shared across agents to guide decisions such as mesh sizing, boundary condition application, solver reuse, and parameter updates during optimization, enabling recovery from common failure modes through targeted regeneration rather than full pipeline restarts. The framework is validated using several three-dimensional structural examples involving static loading and stress- or displacement-constrained optimization, demonstrating that physically consistent finite element models can be generated and design parameters iteratively updated to satisfy prescribed performance targets.