Traffic forecasting has gained more and more interests in both academic and industrial researches. The time series-based models are firstly applied to deal with linear dependency but cannot describe the nonlinear and complex properties of traffic data. Recently, many methods for traffic forecasting based on machine learning and deep learning approaches are proposed. However, these models always encounter the unsolved questions relating to the reliability and the feasibility. Indeed, traffic forecasting is a very challenging task due to the complex spatial correlations in road network, the high-level time dependency and the difficulty of long-term prediction. To address the mentioned challenges, we propose a novel system based on multi-agent systems approach called P-ADRIP (Prediction subsystem-Adaptive multiagent system for DRIving behaviors Prediction) that aims to provide dynamic and real-time traffic prediction. The conducted experiments demonstrate the outstanding performance of ADRIP comparing to the state-of-the-art prediction methods.