Virtual training allows the learning and rehearsal of implicit cues, e.g., trustworthy leading action in an emergency evacuation, that cannot be easily understood through merely reading about situations, while mitigating the danger and expense of live rehearsals. We have focused our efforts on designing social agents that can engage in and help to train humans to generate the trustworthy behaviors that help to ensure a successful evacuation. Drawing upon social science research and using a “role-reversal method,” we successfully constructed agents that can perceive trustworthiness as humans do. The agents first collect human responses to their own nonverbal cues in controlled experimental training scenarios. Using these results, we obtain optimal parameters for nonverbal cues of trustworthiness, and then can use them to guide agents who evaluate human performance in the same training scenarios. The method enables us to convert social psychological findings into computational mechanisms.