Abstract: Unstable user connections and the complex coupling of multi-dimensional system optimization severely constrain the performance of unmanned aerial vehicle (UAV) cluster communication systems in jamming environments. Malicious jammers frequently attack these clusters with jamming signals, severely disrupting the air-to-ground communication links during service provision. Facing this adversarial scenario, a joint optimization method based on multi-agent deep reinforcement learning was proposed to enhance anti-jamming user access. Specifically, a jamming-aware state space, a joint three-dimensional (3D) trajectory and transmit power control action, and a multi-dimensional composite reward function were constructed. This approach achieved the coordinated optimization of UAV spatial deployment and transmit power, successfully maximizing the number of stably connected users meeting the communication quality threshold. Simulation results demonstrate that the proposed algorithm exhibits superior convergence performance and robustness across diverse user distribution scenarios and jamming intensities. Furthermore, under the condition of a jamming power of 25 dBm and a communication quality threshold of 5 dB, the proposed algorithm improves the system average total number of connected users by up to 56% compared with benchmark schemes, thereby effectively enhancing the anti-jamming access capability of UAV swarm communication systems.