This paper describes the theoretical underpinnings and architecture of a new type of learning environment that incorporates features of microworlds and of intelligent tutoring systems. The environment is based on a progression of increasingly sophisticated causal models that simulate domain phenomena, generate explanations, and serve as student models. Constraints on model evolution are discussed in terms of causal consistency and learnability, and a taxonomy of models useful for instruction is outlined. The design principles underlying the creation of one type of causal model (zero-order models for electrical circuit behavior) are given, and possible progressions with respect to model elaboration, order, and perspective are described in the context of presenting a theory of model evolution. Finally, the architecture that enables the pedagogical tools of the intelligent learning environment is described, with an emphasis on the range of instructional interactions and learning strategies that can be supported. Eight figures are included. (51 bibliographic notes/references) (EW)