GeoAgent: a hierarchical LLM-based multi-agent architecture for autonomous spatial analysis

Large language models (LLMs) provide powerful impetus for spatial analysis tasks in GIS. LLM-based agents demonstrate immense potential to revolutionize traditional spatial analysis workflows. However, designing a multi-agent collaborative architecture that can achieve autonomy for complex spatial analysis, while overcoming the limitations of traditional GIS in flexibility and dynamic adaptability, remains a critical challenge. This paper proposes GeoAgent, a hierarchical multi-agent collaborative spatial analysis framework that enhances automation, intelligence, and reliability in complex spatial analysis. Its three-layer architecture comprises: the Planning Layer that parses requirements and formulates execution plans; the Execution Layer that schedules execution agents through a manager agent; and the Review Layer that achieves autonomous verification and optimization. The framework integrates three core toolsets enabling environmental perception and data insight capabilities. Experiments across 147 spatial analysis tasks show GeoAgent achieved 94.56% and 95.24% success rates using GPT-4o and DeepSeek-V3 configurations, respectively. Ablation studies confirmed the critical role of Planning and Review Layers in ensuring workflow integrity and result reliability for complex tasks. GeoAgent contributes a novel paradigm for automated intelligent spatial analysis while validating hierarchical multi-agent collaboration and self-verification mechanisms for geospatial challenges, providing design insights for next-generation autonomous GIS systems.