Virtual training is a relatively novel field in which human beings learn to perform certain tasks by repeatedly executing them in a virtual reality environment. To make such training environments more effective, the agent paradigm has proven to be a useful tool. By conceptualising a training system as a 'virtual tutor', the system may be able to support the trainee in a similar manner as human instructors do, among others by providing personalised feedback and adaptive training material. The current paper describes a project that has as aim to develop a virtual training environment for decision making under stress, targeted at professionals in the public domain. The main contribution is two-fold: first, a global overview of the project is presented. Second, a formal approach is put forward for the design of the training environment, based on agent-based simulation and verification. By generating a computational model of the envisioned system and formally analysing the resulting simulation runs, the behaviour of the system can be studied before its actual implementation, thus providing a method for rapid prototyping.