To address the complexity and real-time challenges of multi-domain collaborative decision-making in large-scale dynamic environment, a swarm cooperation decision-making framework based on large language models, named Courses of Action-Large Language Models (COA-LLM), is proposed. By leveraging prompt engineering, textual conversion of environment situations, and multi-level instruction parsing, a "perception-decision-execution" closed-loop response system is constructed. This framework overcomes the limitations of traditional target allocation methods, which are restricted to handling discrete tasks, and achieves the unification of discrete task allocation and continuous spatial decision-making. Finally, a swarm cooperation scenario is built in a simulation system, where general-purpose large models such as GPT and DeepSeek are employed. Comparative experiments are conducted against algorithms based on optimization and machine learning. The experimental results demonstrate the feasibility and effectiveness of the COA-LLM framework, highlighting its input flexibility and output interpretability, which provides a new paradigm for the development of intelligent command and control systems.