This paper describes the design and empirical validation of three distinct pedagogical agent roles (Expert, Motivator, and Mentor) for college students within the MIMIC (Multiple Intelligent Mentors Instructing Collaboratively) agent-based research environment. The pedagogical agent roles were operationalized by image, animation, affect, voice and script, and were developed in Poser 4 and implemented via Microsoft Agent. Two controlled experiments validated the instantiation of the three roles according to learner perception (N=78) and actual impact on motivation and learning (N=71). The results confirmed that the agent roles were not only perceived by the students to reflect their intended purposes but also led to significant changes in learning and motivation, as designed. Specifically, the Expert agent led to increased information acquisition, the Motivator led to increased self-efficacy, and the Mentor led to overall improved learning and motivation. The implications for intelligent tutoring and multi-agent system design and development is discussed.