An AI agent (or "agentic AI") is a program that pursues goals, uses external tools, and takes autonomous actions, contrasting sharply with "tool AI" like standard chatbots that only answer narrow questions. These agents are typically driven by large language models (LLMs) and can perform multi-step tasks, interact with their environment, and are commonly used for automation—such as booking travel plans from a simple prompt. The concept has deep roots, from Oliver Selfridge's 1958 "Pandemonium" paper to 1990s belief–desire–intention (BDI) models, but modern LLM-based agents accelerated after OpenAI's function-calling API in 2023 and Anthropic's Model Context Protocol (MCP) in 2024, with the term "agentic" popularized by researcher Andrew Ng. Training and evaluation often use simulated environments like Minecraft, while their autonomy is compared to self-driving car levels—most operate at level 2 or 3, with level 5 remaining theoretical. Architecturally, Ken Huang proposed a seven-layer reference model covering foundation models, data operations, frameworks, deployment, evaluation, and security, highlighting the complexity behind these systems. This makes AI agents a rapidly evolving field, moving from simple rule-based logic to sophisticated, goal-oriented digital workers.