Energy- and service-time-intensive IoT applications create major challenges in computation, communication, and resource management. Although traditional cloud systems offer scalability, they often lead to high energy usage and long service delays, making them impractical for tasks that are both delay-sensitive and energy-restricted. Multi-access Edge Computing (MEC) helps reduce these problems by moving computation closer to IoT devices, yet effective task offloading still faces difficulties due to diverse system architectures and dynamic workloads. This paper introduces a Q-Learning-based offloading framework that adaptively allocates computational tasks across mist, edge, and cloud layers while considering system states, task features, and available resources. Simulation outcomes reveal that the proposed method notably decreases energy consumption and service time, while ensuring balanced CPU utilization and outperforming static offloading methods. The findings underline the capability of reinforcement learning to achieve energy-efficient and low-latency IoT services within emerging MEC environments.