The shift of large language models (LLMs) from passive text generators to agentic systems marks a move toward autonomous intelligence. Agentic AI refers to systems that reason, plan, and execute actions in pursuit of complex goals with minimal human intervention. This paper reviews recent research on the architectural foundations, core functional modules, applications, and safety challenges of autonomous LLM systems. Current work emphasizes modular architectures in which an LLM serves as the central reasoning engine, augmented by components for perception, memory, planning, and tool interaction. Multi-agent systems have emerged as a dominant trend, enabling collaborative problem solving through structured orchestration and communication protocols. Although these systems show significant potential in software engineering, manufacturing, and cybersecurity, critical barriers remain in safety, reliability, and ethical governance. This review provides a structured overview of the state of the art and identifies key research directions for the coming years.

Index Terms—agentic AI, autonomous agents, large language models, multi-agent systems, retrieval-augmented generation, tool use, AI safety