What is AI agent?

What it is

A chatbot answers a question and stops. An AI agent keeps going. You give it a goal, such as "find the cheapest flight and put it on my calendar," and it works out the steps, takes them, checks what happened and adjusts until it is done or stuck.

Think of the difference between asking a colleague for directions and asking them to drive you there.

How it works

Most agents run a loop around a language model:

1. Look at the goal and everything that has happened so far.
2. Decide the next action, such as searching the web, running code, reading a file or calling another program.
3. Act through a tool the agent has been given.
4. Read the result and feed it back in as new information.
5. Repeat until the goal is met or a limit stops it.

The model supplies the reasoning, and the tools supply the reach. Without tools, it is just a chatbot. Each trip around the loop is another round of inference, and the history grows each time, which is why agents lean on a large context window and on prompt caching to keep costs sane.

Agents range from simple to ambitious. A basic one might follow a fixed routine with a model filling in the blanks. A more autonomous one picks its own steps, and some can hand parts of the job to helper agents.

Why it matters to you

Agents are where AI moves from drafting to doing. That is useful, and it raises the stakes:

• Mistakes compound. A wrong guess in step two can carry through to step ten. Agents work best on tasks that are easy to check.
• Permissions matter. An agent can only misuse what it can reach. Give it the narrowest access that does the job: one folder, one account, read-only where possible.
• Costs add up. Many loops mean many tokens, so a long task can cost far more than a single chat.
• Untrusted input is a risk. An agent that reads web pages or email can be steered by instructions hidden in them, which is why red teaming agents is a growing practice.

How to start

Pick a task that is boring, repetitive and easy to verify. Let the agent do it with limited access and review the result before anything is sent, paid or deleted. Raise its freedom only as it earns trust.

Related ideas

An agent is not a different kind of model. It is a model plus tools plus a loop. Workflow automation is similar but follows fixed steps, while an agent chooses its own. Many products now call themselves agents, so ask what the system can actually do on its own.