Résumé. Multi agent based simulations (MABS) have been successfully exploited to model complex systems in different areas. Nevertheless a pitfall of MABS is that their complexity increases with the number of agents and the number of different types of behaviours considered in the model. For average and large systems it is impossible to validate the trajectories of single agents in a simulation. The classical validation approaches, where only global indicators are evaluated, are too simplistic to give enough confidence on the simulation's model. It is then necessary to introduce intermediate levels of validation. In this paper we propose the use of data clustering and automated characterization of clusters in order to build, describe and follow the evolution of groups of agents in simulations. These tools provides the modeller with an intermediate point of view on the evolution of the model. Those tools are flexible enough to allow the modeller to define the groups level of abstraction (i.e. the distance between the groups level and the agents level) and the underlying hypotheses of groups formation.