We present a multi-agent ensemble learning approach for supervised learning and evaluate its performance on a task of nonlinear function approximation. This approach relies on a set of learning agents that self-organize according to cooperation rules. We study the properties of this type of system in terms of explainability and interpretability of the prediction process by analyzing the shapes of the agents and their spatial organization. A comparative study on synthetic datasets generated from nonlinear 2D functions is also conducted to evaluate the performance. The results indicate that our multi-agent approach achieves prediction scores similar to state-of-the-art approaches and introduces new properties contributing to address the problematic of explainability in supervised learning.