This paper addresses the issue of adaptive multi-agent systems and their design based on living systems features such as phylogeny and ontogeny. We argue that the evolutionary design of agents behaviors implies several specific features that are missing in classical evolutionary approaches. Therefore we propose a new approach that would be more adequate to MAS, and present a model for building MAS as the result of evolving, interacting, self-organizing agents. We finally mention a use of such an approach for the embodiment of robots.