This paper presents a self-organizing networked controller for command tracking of multi-agent systems. The agents have individual dynamics, are unidirectionally coupled within a basic communication structure and have to follow a command signal. To avoid large control errors and simultaneously reduce the communication effort, the agents request additional information from other agents, if their local control error exceeds a given bound. As a consequence, the communication topology of the overall system is adjusted to the current control task and, as a result of the self-organization, presents an image of the acting command signal at all times. A quantitative bound of the control error is derived in dependence upon the choice of the information activation condition. An algorithm for designing the communication structure and the activation condition for requesting additional information is presented. The approach is demonstrated by its application to a robot formation problem.