What is an AI agent, actually?

The question

What we told them

There is a real definition, and it is shorter than the pitch decks suggest. An agent is an AI model running in a loop, where the model itself decides when the loop stops. That's the whole thing. The autonomy, the price tag, and the risk all follow from that one design choice.

Each pass through the loop is the same four moves:

• Perceive. The system hands the model the current state — the task, everything that has happened so far, any new information.
• Reason. The model decides what to do next.
• Act. The system executes what the model asked for: look up a record, run a query, draft the email.
• Observe. The result gets fed back in, so the next decision is based on what actually happened, not on what the model assumed would happen.

Then it goes around again, until the model decides the job is done. A chatbot never enters that loop — it answers and waits. Traditional automation runs steps somebody scripted in advance. An agent is the only one of the three where the software decides at runtime how many steps the job takes and when it's finished. (Builders do cap the loop with a turn limit, but that's a safety net, not the steering.)

Which hands you a one-question vendor test: who decides when it stops? If a developer scripted every step, it's automation — possibly exactly what you need, usually cheaper and more predictable, just not an agent. If it answers and waits for you, it's a chatbot. Neither answer is disqualifying. Charging agent prices for either one is.