This paper describes a method of multi-agent analysis and design for reactive, real-time information systems, relating to complex and risk-bearing applications. The fundamental principle consists of using a series of models in “cascade” to shift from an abstract represen- tation of the problems to a formal one of the directly programmable agent (in Java for example). The first basic idea is not to have fixed goals or tasks, but rather for them to be gradually released from the analysis of the interactions between the actors (human or artifacts). The second idea, aims at integrating the space-time constraints according to an individual and collective point of view in a concurrent way. The last one, proposes neither to process on a hierarchical basis nor to laminate the final architecture of the interactions between agents but to, on the contrary define the acquaintance rules and their evolution according to the context. This paper details the various stages of this approach and compares them with other current work.