One of the earliest applications of MAS in power systems is for restoration, as described in Nagata and Sasaki . Restoration is needed after a partial or global blackout has occurred. Operators usually employ an OMS to automatically restore power, by sequentially re-energizing all the buses in the grid so as to serve the loads. At the end of the process, the grid is back to its normal operation mode, as before the event. The objective is to serve the maximum of loads connected to the buses. Several constraints have to be respected: the balance between supply and demand, the capacity of each source, and the voltage limits on each bus and branch.
In the proposed approach, two types of agents are used (Fig. 15.27): bus agents (BAG, one for each bus), and a facilitator agent (FAG). Each BAG tries to restore power to the load connected to its bus and can only communicate with its geographical neighbors and with the FAG. Its behavior follows simples rules:
Fig. 15.27 The proposed MAS is made of two types of agents, each with their own roles: bus agents (BAG) and a facilitator agent (FAG) 
• If the available power is not sufficient, the agent negotiates with its neighbors to find another solution and get a higher power to energize its load.
• If no satisfying solution can be found, the agent sheds its loads to its minimum power.
The FAG facilitates the negotiation by classifying buses according to their voltage level, and choosing one BAG at each voltage level to initiate the restoration process. Several restoration processes can be run in parallel at different voltage levels, which accelerates the overall restoration.
In a later paper , the authors proposed another version of this concept, where FAGs are feeder agents and act as managers for the decision-making process. The architecture shown in Fig. 15.27 is replicated at each feeder, and feeder agents can communicate with each other. This enables the system to operate on a much larger scale, with restoration ‘‘groups’’ cooperating with each other. This approach is an improvement compared to traditional OMS, where everything is coordinated by a central controller. The proposed system is much less complex, as the same algorithms can be replicated and cooperate to achieve their objectives, but could be expected to provide results similar to the ones obtained by centralized algorithms.