BIMgent
An autonomous AI agent for architectural building modeling that controls BIM software through its graphical interface—the first computer-use agent applied to the AEC domain.
Definition
BIMgent, presented at the ICML 2025 Workshop on Computer Use Agents by the Technical University of Munich, demonstrates autonomous architectural modeling through graphical user interface control. Rather than calling APIs, BIMgent operates like a human user—seeing the BIM application screen, understanding what's displayed, and executing mouse clicks and keyboard inputs to build the model. Three layers: a Design Layer transforming multimodal inputs (text, images, sketches) into refined floor plans; an Action Planning Layer decomposing modeling procedures into substeps using official software documentation; and an Execution Layer executing specialized workflows with verification. A memory module supports self-reflection and inter-component cooperation. Implemented in Python supporting Vectorworks, BIMgent achieved a 32% success rate on real-world modeling tasks compared to 0% for all baseline models—validating that computer-use agents can meaningfully reduce manual BIM authoring workload.
Examples
BIMgent receiving a floor plan sketch and autonomously placing walls, doors, windows, and slabs in Vectorworks
Action Planning layer decomposing 'create a 3-bedroom apartment' into 47 discrete BIM operations
BIMgent reconstructing an existing building design from photographs without manual modeling
Nomic Use Cases
See how Nomic applies this in production AEC workflows:
Frequently Asked Questions
BIMgent, presented at the ICML 2025 Workshop on Computer Use Agents by the Technical University of Munich, demonstrates autonomous architectural modeling through graphical user interface control. Rather than calling APIs, BIMgent operates like a human user—seeing the BIM application screen, understanding what's displayed, and executing mouse clicks and keyboard inputs to build the model. Three layers: a Design Layer transforming multimodal inputs (text, images, sketches) into refined floor plans; an Action Planning Layer decomposing modeling procedures into substeps using official software documentation; and an Execution Layer executing specialized workflows with verification. A memory module supports self-reflection and inter-component cooperation. Implemented in Python supporting Vectorworks, BIMgent achieved a 32% success rate on real-world modeling tasks compared to 0% for all baseline models—validating that computer-use agents can meaningfully reduce manual BIM authoring workload.
BIMgent receiving a floor plan sketch and autonomously placing walls, doors, windows, and slabs in Vectorworks. Action Planning layer decomposing 'create a 3-bedroom apartment' into 47 discrete BIM operations. BIMgent reconstructing an existing building design from photographs without manual modeling.
Project Research: Instantly access all project-critical information from a single search interface.