This paper illustrates some of the capabilities of previously proposed network control system (NCS) architectures to carry on functioning in the event of faults, without recourse to system reconfiguration. The principle of interaction prediction is used to set up a coordination strategy that encapsulates an ability to withstand or tolerate certain faults, thereby allowing the system to continue functioning. It is also shown that the coordination strategy can be made more effective if a learning agent is allowed to learn the coordination functions. This facilitates the use of different types of agent at the local level, together with recurrent networks and genetic algorithms (GAs) at the coordination level. The experimental test-bed system is a benchmark three-tank system that has some of the main features of an industrial process control system.