Abstract
<title>Abstract</title> <p>The growing sophistication of the contemporary computing environments has heightened the demand of having autonomous intelligent systems that can execute under uncertain and time-varying environments. Specifically, massively scaled cloud environments are dynamic, resource contended, and changing operational limits, whose constant optimization goals cannot ensure consistent performance. Traditional artificial intelligence and reinforcement learning frameworks typically use predefined goals which restricts their capacity to respond to changing needs of the system like being sensitive to latency, cost-efficient, and use of resources. This paper will eliminate this limitation by introducing a self-evolving agentic artificial intelligence system to simulate adaptive cloud scheduling wherein reinforcement learning is joined with meta-learning to facilitate autonomous adaptation of goals. The structure of the proposed approach is the dual-level optimization architecture where action learning layer computes the best scheduling policies, whereas the goal evolution layer updates the objective priorities according to the system feedback. It allows the agent to automatically change its optimization target based on the workload variations and the availability of resources. The framework is tested with real-world workload traces that are the Alibaba Cluster Trace dataset that provides a realistic simulation of cloud behavior with bursty and heterogeneous events of task arrivals as well as resource demands. The experimental findings have shown that the given method shows a consistent improvement compared with the baseline scheduling techniques, such as FCFS, Round Robin, heuristic methods, and regular DQN. Particularly, the model involves a considerable decrease of task latency and operational cost, enhanced resource utilization, throughput, energy efficiency. Those findings prove that empowering intelligent systems to adjust their goals besides optimizing activities is an efficient and scalable answer to dynamic environments. The suggested structure develops a different line of agentic AI systems, where goal evolution and policy learning interact together to attain extended performance in practical settings.</p>