Researchers have presented CADSmith (arXiv:2603.26512), a multi-agent pipeline for generating computer-aided design models from natural language descriptions. The system uses programmatic geometric validation rather than visual feedback to ensure dimensional accuracy in generated 3D models.
Existing text-to-CAD methods face a fundamental precision problem. Single-pass generation produces models without verifying dimensional accuracy, while systems using visual feedback — comparing rendered images to specifications — cannot detect errors that are invisible in 2D projections. For engineering applications, a bracket that appears correct in a rendering but is 2mm off in a critical dimension will not fit its assembly.
CADSmith generates CadQuery code from natural language and subjects it to two nested correction loops. An inner loop resolves code execution errors including syntax issues and runtime failures. An outer loop performs geometric validation, checking that the 3D model satisfies dimensional and spatial constraints extracted from the original specification.
The multi-agent architecture assigns specialized roles: one agent generates initial code, another validates geometric properties, and a third mediates between design intent and physical constraints. This separation mirrors human CAD workflows where designers, engineers, and quality inspectors handle different aspects of the process. The approach moves text-to-CAD closer to engineering-grade utility, where dimensional accuracy is non-negotiable.
