AI agent
An AI agent is a system where a language model pursues a goal by taking actions with tools, deciding its own next step, rather than producing text in response to each prompt.
An AI agent is a system in which a language model works towards a goal by taking actions, rather than by answering one prompt at a time. The model is given tools it can call, some memory of what it has already done, and a loop: choose an action, observe the result, decide what to do next, stop when the goal is met.
What separates an agent from a chatbot is who decides the next step. In a chatbot a person supplies each turn and the model responds. In an agent the model chooses the sequence itself, which means it can do work while nobody is watching, and it means the choices it makes are consequential in a way a wrong answer in a chat is not. An agent that can send email can send the wrong email.
That consequence is why the interesting engineering in an agent is not the model. It is the boundary: which tools exist at all, what limits are enforced in code rather than requested in a prompt, which actions require a person to approve them, and whether every action is recorded with the reason behind it. A prompt asking a model to be careful is not a control. A hard cap on how many messages can be sent in a day is.
In practice
- An agent that reads a codebase, edits files, runs tests, and iterates until they pass.
- An agent that researches prospects, writes to them, and answers replies within a daily cap it cannot exceed.
- Not an agent: a model that drafts an email and stops, waiting for a person to send it.
Questions
What is the difference between an AI agent and a chatbot?
Who decides the next step. A chatbot responds to prompts a person supplies, one turn at a time. An agent picks its own next action in a loop, uses tools to affect things outside the conversation, and can run with nobody present. The practical consequence is that an agent's mistakes are actions rather than sentences, which is why agents need enforced limits and audit trails that a chatbot does not.
What stops an agent doing something harmful?
Only the boundary you build. Instructions in a prompt are a preference, not a control, because the model can misread them and there is nothing to stop it if it does. Real controls are structural: not giving it a tool for the thing it must never do, enforcing caps in code so exceeding them is impossible rather than discouraged, requiring human approval for anything that spends money or is hard to reverse, and logging every action so a wrong one can be found and undone.
Do agents need to be supervised?
They need supervision at the right points rather than continuously, which is the whole design problem. Watching every step removes the reason to use one; watching nothing means discovering mistakes after they have compounded. The workable arrangement is that reversible, bounded work proceeds unattended and anything expensive, public or irreversible waits for a person.
From the blog
Related terms
Last updated