What an AI agent actually is

Strip away the marketing and an AI agent is a small, understandable thing. OpenAI frames it plainly: an agent is a model + tools + instructions. The model does the reasoning, the tools let it affect the world, and the instructions tell it what it's for. That's the whole object.

Anthropic puts the behavior just as plainly: agents are "typically just LLMs using tools based on environmental feedback in a loop." Read that carefully, because it's the honest definition. The model looks at a situation, picks an action, the environment responds, and the model looks again. Nothing mystical — a loop that keeps going until the job is done or you stop it.

Here's the golden rule this whole module is built on, and it starts here: the model is the easy part. It's the piece you don't build. You rent it from an API. Swapping GPT for Claude for Gemini is a config change. The model will happily reason about almost anything you ask. What it won't do on its own is stay on task, stop when it should, avoid destructive actions, or tell you honestly whether it succeeded. All of that lives outside the model.

So when people say "we built an agent," what they actually built is almost never the model. They built the scaffolding around it: the loop that calls the model, the tools it's allowed to touch, the limits on those tools, and the checks that confirm the work. We'll call that scaffolding the harness, and the next four lessons are about it.

A useful mental picture:

instructions ->  [ MODEL ]  -> chosen action -> [ TOOL ] -> result
                     ^                                         |
                     +----------- feedback --------------------+

The arrow that loops back is the agent. A model that answers once and stops is just a chatbot. A model whose output feeds a tool, whose result feeds back into the model, which decides the next step — that's an agent.

Two honest limits to hold from day one. First, the model has no memory between calls except what you put in its context, so "it forgot" usually means "you didn't give it the information." Second, the model cannot actually tell whether its own answer is correct; it predicts plausible text, and plausible is not the same as verified. Both limits are things the harness has to cover, not things a bigger model fixes.

Keep the definition small and the expectations honest. A capable model wrapped in a careless loop is a liability. A modest model wrapped in a careful loop is a tool you can ship.