Data integration, in the context of the web, faces new problems, due in particular to the heterogeneity of sources, to the fragmentation of the information and to the absence of a unique way to structure, and view information. In such areas, the traditional paradigms on which database foundations are based (i.e. client/server architecture, few sources containing large information) have to be overcome by new architectures. In this paper we propose a layered P2P architecture for mediator systems. Peers are information nodes which are coordinated by a multi-agent system in order to allow distributed query processing.