low-Earth-orbit (LEO) satellite constellations，leveraging their characteristics of broad coverage and seamless access，are accelerating the emergence of space-air-ground integrated networks (SAGINs) as a highly promising paradigm for mobile edge computing (MEC).However，existing works have not adequately addressed the critical challenge of task-resource matching in scenarios involving multi-satellite collaborations and dual space-air edges，making efficient offloading of time- and energy-sensitive tasks particularly difficult.This paper formulates a multi-objective joint optimization problem，holistically considering the matching relationship between tasks and multi-dimensional resources，the latency and energy consumption of task processing，and the transmission costs associated with inter-satellite collaborations.A multi-agent reinforcement learning (MARL)-based framework for collaborative multi-task offloading in SAGINs is proposed.This method effectively integrates the cross-domain collaborative decision-making among satellites，unmanned aerial vehicles (UAVs)，and ground stations，along with inter-satellite collaborative decision-making.The experimental results demonstrate that the proposed approach achieves efficient convergence and exhibits significant advantages compared with existing methods.