This paper studies the trajectory tracking control of a class of multi-agent systems, in which, each agent is expressed by a nonlinear discrete-time unknown dynamics and can interact with its neighbors via the history outputs of its neighbors. In order to tackle each unknown dynamics, based on its neighbors' and its own history I/O data and neural network, an approximate model is established by the direct data-driven method. Using the neighbors' history information and the reference trajectory, the decentralized adaptive indirect data-driven control is designed; then, the feedback gain matrix online is designed and adjusted by measured output data and previous estimates. For each agent, this is an adaptive control process of prediction, estimation, and adjustment, which needs to solve some nonlinear optimization problems online, can surmount the negative effects of the modeling errors caused by neural networks, and is the key to making each agent output asymptotically track the given reference trajectory. The convergence analysis shows that the applied method is effective and feasible. INDEX TERMS Multi-agent system, tracking control, indirect data-driven method, adaptive control, nonlinear discrete-time system.