This paper addresses the fault-tolerant control (FTC) problem for multi-agent systems (MASs) subject to concurrent sensor faults and non-affine faults. A hierarchical predefined-time optimal FTC framework is proposed to enhance system reliability and convergence speed. Leveraging hierarchical control architecture, reinforcement learning (RL), and predefined-time stability theory, a synergistic mechanism between virtual and actual control layers is established. Within the virtual layer, a distributed optimal consensus tracking controller is developed. This incorporates an adaptive state observer to estimate unknown nonlinear dynamics and integrates sliding mode surfaces with an actor-critic structure to derive an approximate optimal control policy. This approach simultaneously achieves multi-agent synchronization, optimal control, and energy minimization. The actual control layer synthesizes an adaptive fuzzy predefined-time FTC tracking controller based on optimal trajectories generated by the virtual layer. Utilizing the Lyapunov stability theory, this controller guarantees that tracking errors converge to a predefined-time bounded set within a user-specified time, while concurrently approximating sensor fault parameters and non-affine fault functions. By tuning predefined parameters in the designed controllers, the system achieves adaptive tracking objectives within any desired, user-defined time. Numerical simulations validate the effectiveness of the proposed control strategy.