What turns a chatbot into an AI agent?

A chatbot answers; an agent acts. An AI agent is a program that can pursue a goal, use software or other tools, and take actions with some level of autonomy. A chatbot in its plain form takes a question and returns an answer, and that is the end of it. Wikipedia describes agentic AI as the contrast to that kind of narrow, tool-like use, such as the non-agentic chatbots that were common in 2023.

The idea is older than chatbots, and it is broader than you might think. An intelligent agent is an entity that perceives its environment and takes actions autonomously to achieve goals. By that definition a basic thermostat counts, and so does a human being, or a firm. The thermostat senses the room, has a goal temperature, and acts by switching the heating on or off.

What is new is what sits inside. Modern AI agents are frequently driven by a large language model, but the model alone is not enough. A language model by itself is typically not an autonomous agent, because it cannot interact with dynamic environments, recall past behaviors or plan future actions. So a layer of ordinary software, sometimes called a harness, surrounds the model and manages the prompts, the tools, the memory and the permissions. Engineers at Anthropic put the result simply: an agent is typically just a language model using tools, guided by feedback from the environment, in a loop.

There is no universally agreed definition, so the word is used loosely. The attributes people usually list are pursuing a goal, using tools, changing things in the world and finishing several steps on its own.