We are developing an embedded training system to teach students to use complex software applications. Our tutor reasons strategically about its pedagogical plans. It uses a discourse model to communicate with the student and to guide its strategic decision making. We are developing the tutor using the Collagen platform for collaborative agent management, which provides a plan representation language, plan recognition capabilities, and a discourse model. Our embedded training system uses instrumentation software we have developed to allow Collagen to communicate with the application the student is learning. It also includes a student model to assess the state of the student’s domain knowledge. In this paper we discuss the need for dialogue in a strategic embedded training system. We then describe the system architecture and present an example of a possible training scenario.