WHAT IS AN AI AGENT? A WORKING DEFINITION
A definition narrow enough to be useful: a model that decides its own next action, in a loop, against real tools. If the sequence is fixed, it is a pipeline — and that is usually better.
An AI agent is a language model that chooses its own next action, in a loop, against a set of tools, in pursuit of a goal. The load-bearing word is chooses. If your code decides what happens next and the model only fills in the text, you have a pipeline — and a pipeline is usually the better engineering choice.
The loop is the whole idea#
Strip away the vocabulary and every agent is the same four steps, repeated: the model reads the current state, picks a tool, the tool runs, and the result is appended to the state. Then it decides whether to go again.
That is it. Everything a framework adds — memory, planners, role hierarchies, handoffs — is scaffolding around that loop. Understanding this makes the framework comparisons much easier to read, because you can ask of each one: what does it do to the loop?
What separates an agent from a chatbot#
- It acts. Tools have effects outside the conversation — a query runs, a file is written, an API is called.
- It iterates. One request can produce many steps, and the number is not known in advance.
- It decides. Which tool, with what arguments, and when to stop, are the model's calls rather than yours.
Take away the third and you have an orchestrated workflow. That is not a lesser thing — it is more predictable, cheaper to run and far easier to test. Plenty of production "agents" are workflows, and are better for it.
Why the definition matters commercially#
Agency is what makes these systems useful and what makes them expensive to operate. A model that decides its own steps can also decide wrong ones, in a loop, at whatever your per-token rate is. Every serious design question — budgets, step caps, human approval gates, whether a tool may write as well as read — follows from taking the definition seriously.
It is also where the frameworks genuinely differ. LangGraph makes you state the permitted transitions up front. AutoGen lets structure emerge from conversation. Neither is wrong; they are bets on how much you trust emergence in your particular domain.
The honest test#
Ask what happens on the tenth iteration when the model is confidently wrong. If you cannot answer, you do not have an agent design — you have a demo. The answer usually involves fewer tools, a hard step cap and a human in the loop at the one point where the action is irreversible.