While there has been a significant amount of research on technical issues regarding the development of agent-based learning environments (e.g., see the special issue of Journal of "Interactive Learning Research," 10(3/4)), there is less information regarding cognitive foundations for these environments. The management of control is a prime issue with agent-based computer environments given the relative independence and autonomy of the agent from other system components. This article presents four dimensions of control that should be considered in designing agent-based learning environments, with the Multiple Intelligent Mentors Instructing Collaboratively (MIMIC) system as an example. The first dimension of control involves instantiating the "instructional purpose" of the environment on a constructivist (high learner control) to instructivist (high program/agent control) continuum. The second dimension entails managing "feedback," involving issues of type, timing, amount, explicitness, and potential for learner choice. Third, agent versus learner control is further defined through the desired "relationship" of the learner to agent(s) (e.g., agent as learning companion, agent as mentor, multiple pedagogical agents or agent as personal assistant). Fourth, to be instructionally effective, the agent(s) must assert enough control so that the learner develops "confidence" in the agent(s) in terms of believability, competence, and trust. Overall, an array of possible permutations of system versus learner control must be carefully considered.