The realization of complex distributed applications required in areas such as e-Business, e-Government, and ambient intelligence calls for new software development paradigms, such as the Service Oriented Computing approach, which accommodates for dynamic and adaptive interaction schemata, carried on at a peer-to-peer level. Multi Agent Systems promise to offer the natural architectural solutions to several requirements imposed by such an adaptive approach. Nevertheless, architectural design patterns commonly adopted to model agent interaction schemata do not totally fit the needs implied by highly adaptable service oriented computing scenarios.In this paper, the agent oriented software engineering methodology Tropos is adopted to discuss common architectural agent patterns, and propose architectural improvements to better deal with high degrees of dynamic system adaptability. The novel architectures are illustrated through a case study in the field of ambient intelligence, namely a real Multi Agent System application implementing a pervasive museum visitors guide.