Creation of compelling 3-dimensional, multi-user virtual worlds for education and training applications requires a high degree of realism in the appearance, interaction, and behavior of avatars within the scene. Our goal is to develop and/or adapt existing 3-dimensional technologies to provide training scenarios across the Internet in a form as close as possible to the appearance and interaction expected of live situations with human participants. We have produced a prototype system, JackMOO, which combines Jack, a virtual human system, and LambdaMOO, a multiuser, network-accessible, programmable, interactive server. Jack provides the visual realization of avatars and other objects. LambdaMOO provides the web-accessible communication, programability, and persistent object database. The combined JackMOO allows us to store the richer semantic information necessitated by the scope and range of human actions that an avatar must portray, and to express those actions in the form of imperative sentences. This paper describes JackMOO, its components, and a prototype application with five virtual human agents.