A collaborative net is a sparse and flat organization of agents (decision-makers such as relays, optimization programs and humans). Being flat, collaborative nets can cut across and integrate the many different hierarchic organizations that operate large networks, such as the electric grid, traffic systems, and computer networks. Even when each agent is severely limited in what it can observe and do, there are simple protocols by which it can collaborate (exchange data) with its neighbors, who can collaborate with their neighbors, and so on, to provide all the agents with a common purpose and make it possible for them to achieve global goals. One of these protocols is a distributed version of model-predictive-control (the rolling horizon approach to control). It provides many opportunities for automatic learning. I will describe this protocol, and the opportunities it provides, with the help of some experiments in the real-time control of a small but prototypical network.