We consider two approaches for building component-based systems, which we call respectively architecture-based and architecture-agnostic. The former consists in describing coordination constraints in a purely declarative manner through parameterizable glue operators; it provides higher abstraction level and, consequently, stronger correctness by construction. The latter uses simple fixed coordination primitives, which are spread across component behaviour; it is more error-prone, but allows performance optimisation. We study architecture internalization leading from an architecture-based system to an equivalent architecture-agnostic one, focusing, in particular, on component-based systems described in BIP. BIP uses connectors for hierarchical composition of components. We study connector internalization in three steps. 1) We introduce and study the properties of interaction expressions, which represent the combined information about all the effects of an interaction. We show that they are a very powerful tool for specifying and analyzing structured interaction. 2) We formalize the connector semantics of BIP by using interaction expressions. The formalization proves to be mathematically rigorous and concise. 3) We introduce the T/B component model and provide a semantics preserving translation of BIP into this model. The translation is compositional that is, it preserves the structure of the source models. The results are illustrated by simple examples. A Java implementation is evaluated on two case studies.