A Systematic Review of Agentic AI - Architectures, Evaluation, and Safety

This paper presents a systematic review of agentic AI, focusing on how large language models are engineered into systems that plan, use tools, and interact with real environments. It proposes a practical taxonomy covering interfaces, memory, reasoning/planning, tool use, execution/guardrails, feedback/self-improvement, and multi-agent coordination. The study also categorizes major benchmarks (web, software engineering, and desktop automation) and summarizes evaluation metrics for success, robustness, cost/latency, and safety. Finally, it highlights key risks such as indirect prompt injection and outlines a research agenda for reproducible and trustworthy deployment of autonomous agents.