Architectural Patterns for Large Language Model–Driven Autonomous Agents: A Systematic Framework for Scalable and Robust Agentic Systems

Large Language Models have become efficient reasoning systems that allow autonomous agents who can plan, make decisions, interact with tools, and adapt to the environment. Nevertheless, most deployments are often monolithic in nature and have tightly bound control logic that restricts scalability, robustness, and reusability. This paper proposes a systematic architectural framework of autonomous agents based on LLM, which suggests a taxonomy of reusable patterns that separate reasoning, planning, memory, execution, and self-evaluation functions. The fundamental patterns presented are Planner-Executor to separate the strategic planning and action implementation, Reasoner-Critic to perform iterative self-assessment and refinement of outputs, Memory-Augmented Agent to maintain persistent knowledge management beyond the scope of the situation, Tool-Oriented Agent to combine external capabilities in an organized manner, and Hierarchical Agent Supervisor to coordinate multiple specialized agents. A reference architecture that comprises these patterns gives specific boundaries to the components, which facilitate modularity, fault tolerance, and scalability. Experimental analysis has shown that pattern-based architectures are much more effective at completing tasks, being able to reason accurately, and recover from failures than monolithic agent designs, especially with complex tasks with long horizons. The proposed structural designs offer fundamental infrastructure for the creation of scalable, supportable, and reliable agentic AI infrastructures within various areas of usage.