From a system's perspective as opposed to an individual agent perspective, MAS adaptation is now becoming an important topic, since it can help to obtain expected outcomes under changing circumstances. In this paper, we propose a MAS architecture (2-Lama) to adapt social conventions in dynamic systems. The proposed architecture consist of two layers: the conventional MAS system (we call domain-level) and an additional layer or meta-level in charge of adaptation. A Peer-to-Peer scenario helps us to illustrate our approach. The resulting model changes participant organisation depending on environment and agent changes. Finally, we present some preliminary results of the empirical evaluation of our approach. Categories and Subject Descriptors I.2.11 [Articial Intelligence]: Distributed Articial IntelligenceMultiAgent Systems General Terms Design, Algorithm. Keywords Multiagent Systems, Assistance, Coordination, P2P.