AgentFly is a software prototype providing a distributed architecture for large-scale NAS-wide simulation implemented as a multi-agent system. AgentFly is implemented on top of the Aglobe platform which is both an implementation framework and a runtime engine for custom agents. It was selected over possible alternatives (e.g. JADE) for its outstanding performance and scalability supporting seamless interaction among heterogeneous software, hardware and human actors. AgentFly system has been developed for over five years. It was initially built f or simulation-based validation and comparison of various approaches for autonomous collision avoidance algorithms adopting the free-flight concept. Later, AgentFly has been extended with high-level control algorithms providing tactical control - i.e. coordination of several autonomous unmanned aerial vehicles (UAV). The same agents and algorithms integrated in AgentFly simulation are also deployed on real UAV platforms. Besides this UAV-related application, the U.S. Federal Aviation Administration (FAA) supports the application of the AgentFly system for simulation and evaluation of the future civilian air-traffic management system. AgentFly has been extended with high-fidelity computational models of civilian airplanes and a parallelization concept integrating dynamic load-balancing. The parallelized approach of AgentFly has been validated in simulation using data of a full civilian air-traffic touching NAS. Nowadays, AgentFly is being extended so that it provides a simulation of ATC functions for the NEXTGEN concept validation. There are being integrated ATC & NAS automation agents which are used to simulate human operation in ATM.