Today embedded applications become large scale and strongly constrained. They require a decentralized embedded intelligence generating challenges for embedded systems. A multi-agent approach is well suited to model and design decentralized embedded applications. It is naturally able to take up some of these challenges. But some specific points have to be introduced, enforced or improved in multiagent approaches to reach all features and all requirements. In this article, we present a study of specific activities that can complement multi-agent paradigm in the ”embedded” context.We use our experience with the DIAMOND method to introduce and illustrate these features and activities.