<p>The rapid expansion of latency-sensitive and data-intensive IoT applications has accelerated the adoption of cloud–fog computing as a hybrid computational paradigm. But, the highly dynamic network topology, heterogeneous resource capabilities, and fluctuating user demands create significant challenges in developing an adaptive, delay-efficient, and energy-efficient task scheduling framework. Traditional metaheuristic approaches, such as multi objective Harris Hawks Optimization (MoHHO), provide strong global search capability but face limitations in real-time operation due to their iterative nature. Conversely, Deep Reinforcement Learning (DRL) offers rapid online decision-making but often struggles with convergence and requires large volumes of high-quality training data. To overcome these limitations, this paper proposes a hybrid Multi objective Harris Hawks Optimization and Proximal Policy Optimization (MOHHO-PPO) algorithm for optimal task scheduling in cloud–fog computing environments. In the proposed framework, MOHHO acts as an offline expert optimizer that generates Pareto-optimal scheduling samples across varying workload and network conditions, while the PPO agent is trained to learn and generalize the optimal scheduling policy. During runtime, the trained PPO agent provides low-latency decisions, whereas MOHHO periodically refines the policy through adaptive expert demonstrations to maintain robustness under system variations. The hybrid architecture enables a balance between global exploration, local exploitation, and real-time adaptation. Simulation results demonstrate that MOHHO-PPO significantly improves delay, energy consumption, and network load compared to MoHHO, PPO-only schedulers, MOGWO, and cloud–fog cooperation algorithms. The proposed system reduces transmission delay by up to 32%, energy consumption by 28%, and overall network load by 21%, demonstrating superior adaptability and robustness in dynamic IoT environments.</p>