To fill the existing gap between agent-oriented methodologies and multi-agent system infrastructures, we studied how to build a new methodology that takes into consideration the infrastructures' features. Our aim was, starting from specific requirements, to reuse "fragments" of existing methodologies by composing them, because they are documented, known and tested. To this purpose we exploited the Situational Method Engineering (SME) approach, and used the OMG standard Software Process Engineering Metamodel (SPEM) as a tool for describing the processes. In this paper we present the motivations of our approach, the work we have done to compose the new methodology, and briefly the resulting methodology.