This thesis presents the development and implementation of a system built upon an abstract architecture named ART4FL, designed to address critical problems in traditional federated learning systems. Whilst traditional federated learning approaches work to preserve data privacy, they often rely on centralized servers, introducing a single point-of-failure, which can easily undermine the robustness and effectiveness of the system. To address these limitations, the research done and presented in this thesis integrates a multi-agent system into the federated learning process, enabling autonomous negotiation, training and collaboration among agents without the need for central coordination. The thesis details the process of realizing, implementing, demonstrating and evaluating the ART4FL architecture. The demonstration is conducted through a series of tests, to show core system concepts as well as system scalability, where agents independently train and aggregate models in a decentralized manner. The findings of the research done, indicates that the proposed abstract architecture is feasible with potential applications in data-sensitive fields where attributes like security, robustness and scalability are needed.