This letter studies the consensus tracking control for multi-agent systems (MASs) of general linear dynamics considering heterogeneous constant known input and communication delays under a directed communication graph containing a spanning tree. First, for open-loop stable MASs, a distributed predictive observer is proposed to estimate the consensus tracking error and to construct the control input that does not involve any integral term (which is time-efficient in calculation). Then, using the generalized Nyquist criterion, we derive the conditions for asymptotic convergence of the closed-loop system and show that is delay-independent. Subsequently, another observer is designed that allows the MASs to be open-loop unstable. Next, we use the generalized Nyquist criterion to compute the observer's gain matrix. Towards this end, we choose a specific structure with which the problem boils down to computing a single parameter, herein called the predictive observer parameter. Two algorithms are proposed for choosing this parameter: one for general linear systems and one for monotone systems. To the best of the authors' knowledge, this is the first work for which asymptotic convergence of consensus is proven for general linear MASs with arbitrary heterogeneous delays. Finally, the validity of our results is demonstrated via a vehicle platooning example.