Only 130 of Thousands of AI 'Agent' Vendors Actually Build Real Agents, Gartner Finds

A 2025 Gartner review found that only around 130 vendors out of thousands claiming 'agentic' capability were building systems that genuinely qualify as AI agents. The core technical distinction, drawn by Anthropic, separates workflows — where predefined code controls each step — from true agents, where the AI model itself dynamically decides what actions to take. A real AI agent requires five components: a reasoning model, tools to act on the world, short- and long-term memory, a decision-execute loop, and a hard limit to prevent runaway execution. The widespread misuse of the term 'agent' in marketing has made it nearly impossible to evaluate products without understanding this technical definition. The piece argues that reliability challenges, guardrails, and monitoring requirements all stem directly from the agent's core property of autonomous, variable decision-making.