Large-scale decentralized environments, such as the Internet, achieve advantages in exploiting software agents for applications, thanks to their autonomy in carrying out tasks. In such a scenario, interactions among agents are an issue that must be faced in an appropriate way. In the BRAIN framework interactions among agents are fruitfully modeled on the basis of roles, which define a set of capabilities and expected behavior. This approach achieves several advantages, such as agent-oriented features, separation of concerns and reuse of solutions and experiences. In this paper we propose an interaction infrastructure, called Rolesystem, which relies on the aforementioned role-based interaction model. This infrastructure allows agents to dynamically assume roles and interact accordingly. An application example and comparison with other approaches show the effectiveness of our approach.