AI Agents: Definition, Capabilities, Taxonomy

The term AI agent has become one of the most widely used and least precisely defined concepts in artificial intelligence. As organizations deploy agentic AI systems across enterprise environments, the absence of a shared vocabulary creates confusion among practitioners, inconsistency in governance frameworks, and gaps in security postures. This paper addresses that gap by synthesizing definitions from academic surveys, international standards (NIST, ISO/IEC, IEEE), industry frameworks (CSA, OWASP, Gartner, McKinsey), and major technology organizations (Anthropic, Google, Microsoft, OpenAI, IBM, AWS) to establish a consensus definition of AI agents. It identifies fourteen core capabilities that distinguish agentic systems from conventional AI applications, organized into a layered capability model with a five-level maturity scale, and proposes a multi-dimensional taxonomy for classifying AI agents across six independent axes, complemented by an operational catalog of eleven named agent types with associated security profiles. This paper is the first of a multi-part series; the companion document Agentic Reference Architecture addresses the layered architecture of agentic systems and their integrated security overlay. The goal of this paper is to provide security professionals, enterprise architects, researchers, and policymakers with a rigorous foundation for understanding, characterizing, and governing AI agents.