The rapid growth of IoT ecosystems has intensified the complexity of fog–cloud infrastructures, necessitating adaptive and energy-efficient task offloading strategies. This paper proposes MADRL-MAML, a Multi-Agent Deep Reinforcement Learning framework enhanced with Model-Agnostic Meta-Learning for dynamic fog–cloud resource allocation. The approach integrates curriculum learning, centralized attention-based critics, and KL-divergence regularization to ensure stable convergence and rapid adaptation to unseen workloads. A unified cost-based reward formulation is used, where less negative values indicate better joint optimization of energy, latency, and utilization. MADRL-MAML is benchmarked against six baselines Greedy, Random, Round-Robin, PPO, Federated PPO, and Meta-RL using consistent energy, latency, utilization, and reward metrics. Across these baselines, performance remains similar: energy (3.64–3.71 J), latency (85.4–86.7 ms), and utilization (0.51–0.54). MADRL-MAML achieves substantially better results with a reward of $-21.92 \pm 3.88$, energy 1.16 J, latency 12.80 ms, and utilization 0.39, corresponding to 68\% lower energy and 85\% lower latency than Round-Robin. For unseen workloads characterized by new task sizes, arrival rates, and node heterogeneity, the meta-learned variant (MADRL-MAML-Unseen) achieves a reward of $-6.50 \pm 3.98$, energy 1.14 J, latency 12.76 ms, and utilization 0.73, demonstrating strong zero-shot generalization. Experiments were conducted in a realistic simulated environment with 10 fog and 2 cloud nodes, heterogeneous compute capacities, and Poisson task arrivals. Inference latency remains below 5 ms, confirming real-time applicability. Overall, MADRL-MAML provides a scalable and adaptive solution for energy-efficient and latency-aware orchestration in fog–cloud systems.