Workplace conflicts hold significant importance and can greatly impact the overall dynamics of an organization. Effectively managing these conflicts can contribute to a healthier and more harmonious work environment and potentially leading to increased productivity levels. Consequently, it's imperative to provide training in conflict management skills for addressing and resolving workplace conflicts. One method largely used for social skill training consists in role-playing scenarios. Given the cost of such sessions, virtual agents appear to be an interesting tool to set them up. Today, one of the main limitations of using virtual agents is the lack of believability in the agents' non-verbal behavior, not allowing interaction to be as natural as with humans and limiting user engagement. This PhD aims to automatically generate the socio-affective nonverbal behavior of a virtual agent, that will be integrated into a tool for conflict management training. We present the identified steps for developing this agent and provide an overview of the work completed thus far.