Large Language Models (LLMs) have advanced creative AI agents capable of complex planning and decision-making. Yet, autonomous building construction remains difficult due to challenges in handling diverse structures and inferring precise spatial layouts from visual data. We propose a novel multi-agent framework that leverages Large Multimodal Models (LMMs) to integrate textual and visual inputs for iterative building reproduction. In our system, an Advisor Agent performs high-level visual reasoning and planning, while a Constructor Agent translates this guidance into precise blueprints for execution. Additional innovations include part-based structural modeling, blueprint consistency across cycles, and multi-view perception to handle occlusion. While existing multi-agent frameworks primarily focus on dialogue-based or knowledge-intensive tasks, our approach applies specialized agent collaboration to visually-grounded physical construction, addressing the unique challenges of spatial reasoning and iterative structural refinement. Experiments in Minecraft demonstrate that this multi-agent architecture substantially improves accuracy, flexibility, and error correction compared to single-agent baselines.