<p>The rapid dispersion of internet of things (IoT) applications demands real-time data processing with minimum latency and energy consumption, where cloud is not an optimal choice to deal with such types of applications. This led to the emergence of the edge paradigm, which places storage and computation in the vicinity of end devices, offloading decision plays a crucial role for the effectiveness of quality of service parameters to ensure the latency, cost-effectiveness, with optimal performance. The limited computation capabilities, uncertainty, and dynamic environment are the major issues that need to be addressed. Hence, authors have proposed a deep reinforcement learning-based approach for optimal offloading decision and resource allocation in an edge-cloud collaborative environment for IoT applications. Tasks are classified using the tent chaotic honey badger algorithm. Then, a multi-agent-based optimized residual autoencoder integrated deep reinforcement learning model is proposed for the optimal computation offloading and resource allocation. The residual autoencoder approximates the <i>Q</i>-value in the reinforcement learning process. However, the optimal policy of the RL process is selected using the multi-agent-based genetic fire hawk optimization, which effectively converts states into actions. The experimental results show that the proposed technique achieved improved performance in terms of energy consumption, latency, average reward, and success rate than baseline approaches.</p>