Pervasive computing is characterized by an automatic adaptation of services to the users situation. To achieve this we compose services from atomic services taking ambient information into account. Due to changing availability of services because of e.g. user mobility an initial composition is not sufficient but an adaptation during runtime is also necessary. Within our architecture we manage one or more composite services. Decisions regarding their adaptation are made taking requirements from all composite services into account. This is not easy as each composite service may have different goals. In this paper, we present the results of an analysis of pervasive application scenarios. Based on the analysis we developed a two-level decision system for adaptation decisions: on the composite service level decisions are made using linear programming, on the system level using rule based control. The implementation is integrated in our existing service composition architecture.