Royal FLush is a platform for decentralised federated learning based on multi-agent systems. It uses autonomous agents to coordinate training, exchange model updates and reach consensus without a central server. The system is built using SPADE and XMPP, and allows the deployment of experiments across multiple machines. Each agent follows finite state machines that guide its behaviour through training, communication and consensus stages. The platform includes observer agents for passive logging, enabling detailed tracking of performance over time. To test the system, we implement PACoL, an asynchronous consensus algorithm, and run a series of experiments using different topologies, data distributions and numbers of agents. The results show that the approach remains effective in varied conditions and supports reliable learning in a fully decentralised environment.