Geospatial multi-agent collaboration is a new intelligent collaborative paradigm centered around geospatial large models, where multiple agents collaborate, self-organize, self-learn, and self-adapt through continuous optimization and iteration. Based on an analysis of the current development status of geospatial multi-agent collaboration, this study explores its essence and characteristics. It also presents a five-layer system architecture that comprises physical perception, data governance, intelligent engine, multi-agent, and business application layers, as well as three collaboration modes: horizontal, vertical, and hybrid collaboration. Moreover, the study summarizes key technologies such as multimodal data perception and representation, deep integration and understanding of geographical knowledge, intelligent utilization of specialized tools and quality control, as well as collaborative scheduling and self-optimization of multiple agents. Demonstration applications for disaster emergency support service scenarios are also carried out. This study aims to lay a foundation for constructing intelligent systems with capabilities such as task self-planning, tool self-invocation, result self-evaluation, and memory self-evolution. This will advance geospatial applications from static analysis to dynamic intelligent response, providing next-generation decision support for fields such as natural resource monitoring, emergency management, and smart cities.