AI agent
Also known as: agent
An agent is a model wrapped in a loop that lets it do things, not just say things.
A plain chat is one shot: you ask → it answers → done. An agent runs a cycle:
think → act (use a tool) → observe the result → think again → … → done
Typical tools in the loop: running shell commands, reading and writing files, searching the web, calling APIs, opening a browser. After each action, the model reads what happened and decides the next step — often dozens of steps before it comes back to you.
What separates an agent from a chatbot:
| Chatbot | Agent | |
|---|---|---|
| Output | text | text + side effects (files, commands, emails) |
| Steps | 1 | many, self-directed |
| Failure mode | a wrong answer | a wrong action (needs guardrails) |
| You watch | the answer | the loop, step by step |
Because agents act, they need supervision: permissions (what it may touch), a budget (how long it may run), and a human "stop" at the door of dangerous actions. The loop is the power — and the risk.