Distribution systems are now fundamentally changing. This is due to the fact that a lot of small size dispersed generation units (DG) are expected to be installed at the distribution level, that is, in the medium voltage (MV) and low voltage (LV) networks [1.1]. The problems arise from the current design features of distribution networks, which were not supposed to include energy sources. At this level, until now, the energy is bought from the wholesale market, transported through a transmission system and delivered to the customers through a distribution system. This primary objective is compatible with the arborescent operational network configuration, the conditions of protection and security assessment design, the wire characteristics, and also the type of customer service. Since the distribution grid is used strictly in open-loop mode and has no energy sources, the currents are unidirectional and consequently the top-down centralized management model is the strait-forward choice. With the introduction of dispersed energy resources, the network may need reinforcement and the grid protection schemes should be adapted to bidirectional energy flows [1.1, 1.2]. In this case, the amount of information to be treated centrally would grow considerably [1.3], due to the number of generation equipments inserted into the grid. Large amounts of data would need to be collected, treated simultaneously and provided quickly for further processing. The presence of stochastic energy sources and the reliance on the communication system could compromise the centralized grid operation and the security of supply. In this respect, the decentralized approach is adopted in this study to both energy management planning (day-ahead) and on-line operation. The decentralized energy management system is built using a multi-agent approach. This approach allows for the simultaneous integration of competition and collaboration among the actors. It is characterized by a high flexibility and robustness, while using minimum data exchange. Since generation sources at the MV/LV level are of relatively small capacity, their generation is normally consumed locally. This specificity and the arborescent topology (feeders with laterals) tend to create zones containing some local generation corresponding to local energy needs. The multi-agent approach is thus applied to clusters of active distribution networks at two levels: intrazonal and interzonal, based on an original definition of zones. This framework offers a partial autonomy to each cluster and allows for parallel optimization and for cooperation among several independent entities. The decentralized approach was successfully applied to several distribution network problems, such as: unit commitment and dispatch, voltage profile control and supply restoration after a blackout. It seems however evident from this study that an appropriate centralization degree should be preserved in order to maintain an adequate system operation.