To address the challenge that static decision authority allocation in manned-unmanned teaming (MUM-T) cannot adapt to dynamic battlefield situations, a Hierarchical Adaptive authority allocation algorithm based on Reinforcement learning (HAR) is proposed. The algorithm adopts a two-layer architecture: the upper permission allocation network outputs three-level permission grades based on situational factors, while the lower action policy network executes tasks under permission constraints. Permission switching stability is ensured through danger score bias and exponential moving average smoothing mechanisms. A cognitive load index is introduced to quantify human-machine collaboration effectiveness, and a three-stage progressive curriculum learning framework is employed to enhance generalization in complex tasks. Simulation results demonstrate that HAR achieves a 61% task success rate in difficult tasks, outperforming MAPPO (58%) and other baseline algorithms. The high-permission usage ratio adaptively increases from 14.87% in simple tasks to 69.82% in difficult tasks, with a variation of 54.95 percentage points. The cognitive load variation reaches 0.45, significantly outperforming comparison algorithms, validating that the proposed method effectively achieves dynamic matching between decision authority and task difficulty.