This chapter presents the results of a 10 years research focused on building intelligent tutoring systems (ITS) that aims to represent student’s knowledge without classifying it with predefined stereotypes. We adopted the mental states approach as an alternative to achieve this goal. By modeling students using a multiagent BDI (belief desire intention) architecture, the student knowledge is considered the way it is. The results of the experiment, conducted with a multiagent learning environment to teach ecology for children, reinforced our beliefs about this type of modeling students; however, it points out the need of better mediation actions to guide students interaction. Searching to improve mediation, we decided to experiment with a novel approach in ITS research: to include the teacher as an end user. The results of this ongoing work are also presented in this chapter.