The purpose of this chapter is to discuss the use of multi-agent systems to develop virtual reality training systems. We first review these systems and then investigate the architectures used. We demonstrate an example of our own (RiskMan) and then discuss the advantages and drawbacks of using multi-agent agent approaches in the development of virtual reality training systems. The chapter describes the system architecture of a multi-agent system for risk management (RiskMan) to help train police officers to handle high-risk situations. RiskMan has been developed using a high-level scripting language of a game engine, Unreal Tournament 2004. The major modules are a scenario-based expert system, a narrative engine, a game engine, and a graphics engine. The system integrates a simulation agent, trainee agent, communication agent, interface agent, and scripted agents communicating using games technology.