The rapid growth of vehicular technologies has led to the emergence of the Internet of Vehicles (IoV), which connects vehicles, infrastructures, and other entities. Vehicular Edge Computing (VEC) facilitates low-latency data processing for IoV applications but faces challenges due to task, resource, and scenario heterogeneity. This paper proposes a Task Feature-Aware Multi-Agent Proximal Policy Optimization Algorithm (TF-MAPPO) to address these challenges by integrating task feature prediction with adaptive offloading and resource allocation. By capturing task interdependencies and predicting future tasks, TF-MAPPO dynamically optimizes offloading decisions, improving computational efficiency. Simulation results demonstrate that the proposed algorithm significantly reduces computation cost, delay, and energy consumption compared to baseline methods, enhancing overall system performance.