One of the key problems in the study of multi agent systems in which the agents exhibit continuous behavior is the automatic recognition and analysis of intentional activities based on observable behavior. Such an analysis requires software systems to structure motions into episodes that are meaningful in the application domains, to acquire and maintain models of the activities, and to use such models to reason about multi agent system behavior. In the research sketched in this paper we study the acquisition of episode and motion models for football games. We show that these models allow for the realization of impressive application systems including interactive football TV and agent systems that assist coaches in analyzing their teams.