In this paper a decentralized controller-observer scheme for a multi-agent system is presented. The key idea is to develop, for each agent, an observer of the collective system's state and a motion controller. The observer is updated using only information from the agent itself and from its neighbors; the motion controller is designed in order to allow the team's weighted centroid to track an assigned time-varying reference. Convergence of the overall scheme is proven for directed and undirected communication graphs; moreover the extensions to the case of switching communication topologies and to the presence of saturation in the control input are discussed. Finally, numerical simulations are illustrated to validate the approach.