What Is an AI Agent?

A plain chatbot takes one input and produces one output. An agent runs a loop: the model decides on an action (say, "search for X"), that action actually executes, the result comes back into the model's context, and the model decides on the next action — repeating until it decides the task is done.

The mechanism that makes this possible is usually called "tool use" or "function calling": the model is given a list of available actions with a description of what each one does, and it outputs a structured request to call one, instead of a plain-text answer.

A single-agent setup uses one model with a set of tools to complete a task end to end — useful when the task, however many steps it takes, stays within one coherent line of reasoning.

A multi-agent system splits the work across several model instances with different roles (a "researcher," a "writer," a "reviewer"), each with its own tools and context, coordinating through a shared plan or hand-offs — useful when a task genuinely benefits from separate, specialized passes rather than one model trying to hold every role at once.

The practical risk with agents isn't a rogue AI taking over — it's mundane: an agent that misreads a tool's output, or acts on a subtly wrong assumption, can carry that error into several subsequent steps before anything catches it, since each step trusts the last one's result.

That's why real agent deployments almost always include guardrails: a maximum number of steps, a human approval step before anything destructive or costly, and logging of every action taken — the same caution you'd want around any automated system that can actually do things, not just say things.