ABSTRACT The development of vehicle electrification and connectivity technologies is giving rise to new mixed and power-heterogeneous traffic flows, comprising connected autonomous vehicles (CAVs) and regular human-driven vehicles (RVs), fuel vehicles (FVs) and electric vehicles (EVs). The diverse power characteristics and complex control relationships among these vehicles present significant challenges for ecological control at signalized intersections. To address these issues, this study proposes an eco-cooperative adaptive cruise control framework based on multi-agent reinforcement learning (M*-ECACC). Initially, ecological control units and agent division method are developed to adaptively divide traffic flow into agent units according to the guidance control relation among vehicles. The single-agent speed profile planning module and multi-agent cooperative control module are designed and integrated into M*-ECACC framework. Through distributed computation and centralized optimization, a global ecological control solution with efficiency, safety and energy-saving objectives is generated and implemented for the mixed and power-heterogeneous traffic flow at intersections. Moreover, a process-adaptive multi-agent twin delayed deep deterministic policy gradient network MATD3*, is introduced to improve the accuracy and effectiveness of cooperative control solution. Extensive comparative experiments demonstrate that the proposed M*-ECACC method achieves stable and superior ecological control performance under various mixed and power-heterogeneous traffic flow conditions.