This study addresses the insufficiency of centralized management for high-density autonomous aircraft traffic by proposing a decentralized approach using multi-agent reinforcement learning. The researchers extend this MARL framework to manage traffic flow within complex air corridor networks featuring merges and splits. Policies trained in single-corridor settings are tested on increasingly complex multi-corridor scenarios in a zero-shot manner without retraining. Experimental results show that learned behaviors transfer effectively across varying traffic densities, network geometries, and heterogeneous vehicle performances. The evaluation measures system-level performance through conformance to boundaries, completion rates, average speeds, distance traveled, and inter-aircraft separation. Despite requiring only locally coordinated entry, traversal, and exit behaviors, the collective actions produce desirable traffic flows throughout the corridor network.