Cyber-physical production systems are transforming traditional, hierarchical control structures into distributed ones in which the elements offer and consume services with a high degree of autonomy. This paper proposes an agent-based approach to distributed control for production environments in which the agents are only able to interact with a part of the whole system. It hypothesises that the performance of the agent network can be improved through learning and communication. A description of the approach is presented and illustrated with a simulation case of distributed control for an industrial compressed-air system.