To enhance the scheduling efficiency of imaging satellites for regular emergency scenarios (e.g., disaster warning, intelligence acquisition) and address the performance limitations of fixed scheduling algorithms in complex, dynamic situations and the lag associated with manual algorithm design, this study proposes a large language model (LLM)-driven AI agent collaboration framework. The framework decomposes the complex scheduling problem into two subproblems: task allocation and single-satellite planning. The algorithm design process is further divided into two phases: initial algorithm generation and evolutionary algorithm optimization. The resulting executable meta-tasks are autonomously completed by multiple agents through collaborative communication to design the optimal algorithm. The entire framework operates fully automatically and intelligently, supporting offline training and seamless online deployment to ensure real-time response capabilities for emergency tasks. Experimental results demonstrate that algorithms automatically generated by this framework significantly outperform those designed by human experts in both solution quality and computational efficiency.