Most algorithms for learning and pattern discovery in data assume that all the needed data is available on one computer at a single site. This assumption does not hold in situations where a number of independent databases reside on different nodes of a network. These databases cannot be moved to a common shared site due to size, security, privacy, legal, and data-ownership concerns but all of them together constitute the dataset in which patterns must be discovered. These databases, however, may be made accessible for certain types of queries and all such communications for a database may be channeled through an intelligent agent interface. In this paper we show how a decision-tree induction algorithm may be adapted for such situations and implemented in terms of communications among the interface agents.