To enhance the intelligent decision-support capability of command information systems (CIS), this paper proposes employing a data middle platform as a "data foundation" underpinning the CIS. However, the two differ markedly in their modes of application: the CIS centers on a continuous, "linear" command-and-control process, whereas the data middle platform operates on discrete, "point-like" data services as its basic units, making active integration between them difficult. Accordingly, we analyze the characteristics of and differences between their application modes and propose AutoInvoke, an application-integration framework based on large language model (LLM) agents. Upward, it performs intent understanding and task decomposition for the complex tasks of the CIS; downward, it invokes the various tools of the data middle platform to accomplish these tasks. In implementation, to improve the accuracy of task decomposition, a constraint-relationship graph and a domain knowledge graph are constructed; to improve the precision of tool matching, a model matrix is built on the data-middle-platform side to provide unified invocation interfaces, enabling the CIS to invoke the platform''s tools dynamically and intelligently according to task requirements. Finally, a rescue mission is used as a case study to empirically validate the effectiveness of the application integration and of AutoInvoke.