An AI agent is a system built around a language model that can do more than answer: it can decide what to do next, call tools or APIs, observe the result, and continue until a goal is met. The model provides the reasoning; the surrounding software provides the actions and the guardrails.
Agents are powerful precisely because they act against real systems — and that is also why they are risky to deploy carelessly. An agent that can send an email, move money or change infrastructure needs the same controls any automated actor does: scoped permissions, an audit trail, human approval on consequential steps, and a rollback path. The engineering around the model matters more than the model.
Related terms