A reliability-based optimization framework is introduced and used to design filament-wound cylindrical shells with variable angle tow. Seven design cases are investigated to enable a comparison between constant-stiffness and variable angle tow designs, also considering effects of thickness variation created due to overlapping tow paths, determined using the kinematics of the filament winding manufacturing process. The uncertainty in the winding angle is considered in the optimization by means of metamodels constructed using the Kriging method. Moving search windows are incorporated into the Kriging metamodel to accelerate its convergence by reducing the number of training iterations. The results prove the efficacy of the proposed framework and clearly demonstrate the advantage of variable-stiffness designs over conventional ones for achieving a maximum load carrying capacity, while keeping the robustness of the design towards manufacturing uncertainties.