Mass evacuation of people in buildings during emergencies is still a burning question regarding safety and reliability. Due to lack of information, crowd prefers to evacuate in the form of random groups. This random grouping increases the possibility of panic among evacuees due to human behavioral factors like herding and stampeding. As a result, congestion may occur resulting in unnecessary casualties. For this purpose, we propose a multi-agent evacuation architecture, that not only collects all the information from evacuees and events occurring in the building on a real-time basis but also provides the evacuation routes to evacuees. In this paper, we discuss a group formation module of our proposed architecture and use an example as a test case to check the functionality of our proposed group formation approach.