부동산 도메인에서의 에이전트 아키텍처 성능 비교: 멀티 에이전트 워크플로우를 중심으로

Traditional AI research in real estate has focused on single-agent architectures for singular tasks like price prediction. However, real-world decision-making demands comprehensive analysis of heterogeneous information, a task where existing systems show clear limitations. To address this, this paper designs and evaluates several multi-agent workflow architectures that mimic human collaboration, conducting a rigorous quantitative and qualitative comparative analysis against a baseline single-agent system. This study evaluates architectures including a Hierarchical system with a Supervisor Agent and a novel Decentralized (Peer-to-Peer) cooperative system. To objectively verify their utility, we constructed a realistic performance evaluation dataset with queries of varying difficulty, incorporating actual apartment meeting minutes and user reviews to enhance realism. This dataset enables a nuanced assessment of each architecture’s ability to handle tasks from simple information retrieval to complex analytical reasoning. Experimental results demonstrate a clear performance hierarchy. The Decentralized multi-agent system significantly outperformed all other architectures, particularly in the qualitative evaluation of answer completeness and reliability for complex queries. While the single-agent and hierarchical systems achieved high task completion rates, they often provided superficial or factually weak answers. In contrast, the Decentralized system excelled at synthesizing complex information, proving the superior practical applicability of specific multi-agent designs in multifaceted domains like real estate. This provides a crucial foundation for the future development of advanced decision-support systems.