In this paper, we present a mechanism called DWRED (Dynamic Weighted RED) that uses communicating agents in a Multi-Agents Systems to enhance the basic Random Early Detection (RED) congestion management algorithm. Our proposition provides different levels of service for multiple classes of traffic, and adds cooperation between the network nodes that allows dynamic modification of the running algorithm’s parameters in response to different network congestion states. We show preliminary results of a simple DWRED prototype implemented using the Multi-Agents Systems platform MADKIT.