What is an AI agent?
Give a language model a tool and a goal, and it doesn’t just answer a question — it acts. An AI agent is a model hooked up to software that can search the web, run code, call APIs, or perform other concrete tasks, and it decides which tool to use, in what order, until the job is done. Unlike a single-turn chatbot that replies and waits for your next prompt, an agent orchestrates a sequence of steps on its own. The core mechanism hasn’t changed: at each decision point, the model is still running inference on everything it has seen so far, predicting the most useful next move. That shift from answering to doing is powerful — and it raises the stakes when the model is wrong.
How an AI agent works
An agent starts with a goal written in natural language, for example “book a flight to Berlin under €200 that departs after 3pm.” The model breaks this into a plan or simply begins acting. It might first search for available flights, read the results, then use a calculator tool to convert currency or compare prices. Each tool call returns new information that gets added to the model’s context window. The model then reads the updated context and decides the next step — maybe refine the search, maybe book the flight. This loop repeats until the goal is met or the agent decides it cannot proceed.
The decision at each step is still inference: the model is not “thinking” in a new way. It is doing what it always does — predicting the most probable next token — but now the tokens it generates can include special commands that trigger tool calls. The tools themselves are external; the agent just formats instructions for them and waits for the results.
What makes it different from a chatbot
A chatbot processes a single turn. You ask a question, it answers. If you want a follow-up, you type another message. The interaction is linear and you do the steering.
An agent, by contrast, handles multi-step tasks without you intervening. You give it one instruction and it orchestrates the rest. It might search several sources, cross-check facts, write and run a script, then compose a final answer. The conversation feels more like delegating to a capable assistant than chatting with a search engine. But that delegation also means you are not watching each intermediate step — so when the agent makes a mistake, it might compound before you notice.
Why autonomy raises the stakes
When a single-turn chatbot gives a wrong answer, the error is contained. You see it, correct it, and move on. When an agent acts on a wrong inference — say it misreads a flight time and books a ticket for the wrong date — the action is already taken. The agent’s ability to call real-world services means errors can have real-world consequences.
The same model fallibility applies. An agent does not know when it is uncertain; it just generates the next action. A model hallucinating a fact in a chatbot reply is bad. That same hallucination fed into a tool call that generates an invoice or cancels a subscription is worse. Designers add safeguards — human-in-the-loop approvals, step limits, confidence thresholds — but the fundamental risk remains: an agent that acts autonomously can act on bad reasoning. Understanding that the model is still just running inference, step by step, helps explain both the power and the danger. The agent is not smarter; it is more connected.