TO-Agents: Multi-Agent AI Lets Designers Describe Form, Get 3D Topology

TO-Agents, presented by researchers Isabella A. Stewart, Hongrui Chen, and Faez Ahmed, tackles a core engineering challenge: translating qualitative design preferences (e.g., "aesthetically tree-like") into algorithmic topology optimization parameters. The framework employs a multi-agent pipeline: a natural language problem description is converted into validated solver inputs, a topology optimization solver runs, and the resulting 3D shape is rendered. Multi-view vision-language reasoning then allows an independent judge agent to critique the result and revise solver parameters for the next iteration. This cycle repeats four times, enabling the system to explore design space while recovering from poor revisions.

In tests on a cantilever beam and a phone-stand product design, TO-Agents produced at least one preference-aligned design in 60% of trials, outperforming an ablated pipeline lacking visual/historical feedback by up to 6x. The system also includes a manufacturing agent that post-processes top designs for additive manufacturing, creating an end-to-end intent-to-prototype workflow. The authors identify failure modes like overshooting and misplaced tools, underlining that while agentic topology optimization can shift engineers from low-level parameter tuning to higher-level specification, robust safeguards are needed for autonomous design.