This paper introduces an innovative cross-modal spatial layout generation approach that leverages AI-agents to seamlessly integrate graphs, textual descriptions, and geometric boundary constraints for the creation of 3D room layouts. The proposed method utilizes a unique agent-based framework that incorporates large language models (LLM) and graph neural networks (GNN) to process and fuse multimodal inputs, allowing for a more comprehensive and flexible design process. By combining textual descriptions with boundary constraints into room feature embedding, the method enhances the semantic consistency and practicality of the generated layouts. Compared to the previous single-modal generative models, the experimental results demonstrate the method’s effectiveness in accurately reconstructing room layouts and adapting to user-defined changes, showcasing its potential to revolutionize the field of architectural design by enabling the efficient integration of AI-agents and cross-modal data processing. Overall, this paper presents a significant step forward in the development of intelligent, adaptable, and user-centric tools for 3D spatial layout generation.