AgentsCAD uses multi-agent LLMs to auto-fix 3D print geometry defects

3D printing with Fused Deposition Modeling (FDM) often requires manual tweaks—like adding supports or adjusting angles—to avoid print failures. Current slicers can spot defects such as steep overhangs but can't modify the underlying 3D model. A new preprint from researchers introduces AgentsCAD, a multi-agent system that bridges the gap between raw CAD geometry and LLM reasoning. The workflow starts by parsing a STEP file (a standard CAD format). Agents then build a face-adjacency topology graph to identify overhangs exceeding a 45° threshold, and optionally inject semantic feature labels using a GraphSAGE model trained on the MFCAD++ dataset (59,665 parts). A Claude Sonnet design-reasoning agent then recommends specific modifications—reorientations, fillets, chamfers, and similar fixes—while a GPT-4o vision-language verifier inspects rendered views to ensure geometric integrity. The final output is a modified STEP file and a detailed human-readable report.

The system's test on a birdhouse model demonstrated correct diagnosis of overhangs, selection of appropriate mitigation strategies, and generation of physically valid corrections. This work partially solves the geometry-to-language translation problem central to LLM-driven CAD modification, enabling automated Design for Additive Manufacturing (DFAM) without manual CAD expertise. By combining geometric feature recognition with conversational LLM agents, AgentsCAD could streamline the iterative design-to-print loop for engineers and hobbyists alike. The code and paper are available on arXiv (2607.02448).