Developing context-aware applications needs facilities for recognizing context, reasoning on it and adapting accordingly. In this paper, we propose a context-based multi-agent architecture consisting of context aware agents able to learn how to distinguish relevant from non relevant context and to make appropriate decisions based on it. This multi-agent system interacts with a context manager layer, based on an ontological representation of context, which is able to answer context-related queries. The use of this architecture is illustrated on a test MAS for agenda management, using the JADE-LEAP platform on PCs and PDAs.