We are exploring the requirements for and prototypes of intelligent opponents for a virtual reality environment in which U.S. Marines can train to clear buildings of enemy soldiers. This virtual reality trainer is one of the components of the Virtual Training and Environments (VIRTE) Program of the Office of Naval Research. Autonomous, intelligent computer generated forces, such as TacAir-Soar, have been shown to be effective in vehicle-level entity domains. Creating models of individual humans engaged in combat with direct fire weapons presents a host of new challenges. One challenge for the current project is that intelligent agent development is concurrent with development of the simulation technology by other researchers, meaning the final virtual environment is not yet available for the agents. Our solution is to use a commercially available, first-person perspective shooter game (Unreal Tournament) as a prototyping environment. We describe requirements for intelligent opponents, our approach and the current system architecture. To date, we have developed agents that navigate through buildings and defend themselves. Creating much more realistic behaviors is a larger challenge and we introduce a number of potential research questions that could, if pursued, increase the realism of the VIRTE agents.