One of the main advancements in distributed learning may be the idea behind Google’s Federated Learning (FL) algorithm. It trains copies of artificial neural networks (ANN) in a distributed way and recombines the weights and biases obtained in a central server. Each unit maintains the privacy of the information since the training datasets are not shared. This idea perfectly fits a Multi-Agent System, where the units learning and sharing the model are agents. FL is a centralized approach, where a server is in charge of receiving, averaging and distributing back the models to the different units making the learning process. In this work, we propose a truly distributed learning process where all the agents have the same role in the system. We suggest using a consensus-based learning algorithm that we call Co-Learning. This process uses a consensus process to share the ANN models each agent learns using its private data and calculates the aggregated model. Co-Learning, as a consensus-based algorithm, calculates the average of the ANN models shared by the agents with their local neighbors. This iterative process converges to the averaged ANN model as a central server does. Apart from the definition of the Co-Learning algorithm, the paper presents its integration in SPADE agents, along with a framework called FIVE allowing to develop Intelligent Virtual Environments for SPADE agents. This framework has been used to test the execution of SPADE agents using Co-Learning algorithm in a simulation of an orange orchard field.