Cyber-attacks are increasing too fast and becoming more complex for traditional security systems to handle. Some attacks like APTs and zero-day threats will remain hidden for a longer time and hence needs faster and smart protection methods. This paper will focus on using autonomous AI agents that can detect and respond to threats in real time. This system architecture will have agents that will monitor the system, use machine learning for analysing, and take action on its own. The important components are distributed sensors, anomaly based detection systems, decision making units and automatic responders. The agents are trained with online learning and reinforcement learning methods to find and respond to threats in an efficient way. This paper also focuses on attacks that are adversarial in nature and the need for stronger defence. Some points are discussed on scalability, safety and continuous monitoring.