Agent based models (ABMs) can be used to support the assessment of water management policies impact on the resource and its users. ABMs are then to explore the link between collective rules and individual behaviours, and more globally, the influence of constraints belonging to some organization level over dynamics belonging to another one. Our objective is to test how a specific multi-agent architecture which explicitly handles dynamic groups and roles can tackle complexity and enhance models representation power and modularity We experiment our approach over a toy situation in order to properly address conceptual and implementation issues. Organization levels are modelled as groups, functions of agents in these levels are modelled as roles. Roles represent the mechanical aspects of an agent implication within a group : collective rules awareness on the one hand, archetypal organizational behaviors on the other hand. Agents express their individuality choosing or parameterizing their roles. Playing roles in different groups, they are the interaction vectors between the organization levels they participate in. This entails to separate reactive behavioural patterns on the role side and cognitive ones on the agent side.