Machine-to-Machine (M2M) systems compose smart networks of multiple devices sensing or acting in the physical world, and interacting together to provide data to value-added services. M2M infrastructures aiming at being deployed city wide have to deal with the increasing number of services and applications they have to support. A dynamic sharing of M2M infrastructure between applications is thus a strong requirement. In this paper, we propose to use multi-agent abstractions and related programming languages to define an adaptive and agile layer to govern M2M infrastructure to support different applications.