In order to handle complex situations, an autonomous agent needs multiple components ranging from simple input/output modules to sophisticated AI techniques. Integrating a high number of heterogeneous components is a non-trivial task and this paper discusses the use of a symbolic layer to address this issue. After an overview of existing techniques and their limitations this paper proposes a new approach through a generalized hyper-graph model in which the interaction of different components is modeled through a triggering mechanism based on patterns. Finally, the paper shows how a flexible symbolic middleware can be built and a few examples are presented.