This paper considers control of networks modeled by a directional graph. We propose a leader-follower paradigm for coordinating local feedback controllers in a large network of interacting components, where most components - the followers - can handle local load without problems, but where a few heavily loaded components - the leaders - are close to saturation. The leader-follower paradigm partitions the network in subsystems, with one leader in each subsystem. Global performance can be improved if followers control their output variables so that the resulting input variables for their leader improve the performance at their leader. The leader-follower approach is explained using as a case study the coordination of the switching times of traffic lights in an urban traffic network. The local feedback controller at a leader uses local traffic data only, selfishly minimizing local queues by selecting the switching times of its traffic lights. Local controllers at follower intersections minimize a local delay, but also take into account specifications generated by their leader. Simulations show that this approach can improve system performance. The flexible feedback nature of this leader-follower control paradigm in one subsystem of the overall network makes it a good candidate as a building block for a scalable hierarchical controller for the complete network.