What Is an AI Agent, Really?
The term gets used loosely enough that it's worth defining plainly, once, before anything else on this site assumes you already know it. This page does exactly that — and nothing more.
An AI agent is software that reads information, decides what to do with it within rules it was given, and then actually does it — inside a real system, not just in a reply. That's the whole definition. Everything else people mean by "agent" is a variation on those three parts: it reads, it decides, it acts.
Why the Word Gets Confusing
"Agent" gets attached to almost anything with an AI model behind it lately, which makes the word nearly useless on its own. A search box with an AI-written summary on top gets called an agent. So does a script that runs on a schedule with no AI in it at all, if the marketing needed the word. Neither one is what this page means.
The useful test isn't whether something uses AI, or whether it talks in sentences. It's whether it takes an action that changes something — a record gets updated, a message gets sent, a document gets filed — based on a judgement it made, rather than just producing text and stopping there.
What This Page Is Not
This page is a definition, deliberately kept narrow. If you're trying to work out specifically how an agent differs from a chatbot, that comparison lives on AI Agent vs. Chatbot. If you want the mechanics of how the reading-deciding-acting loop actually works under the hood, that's How AI Agents Work. Both build on the definition here rather than repeating it.
What Makes Something an Agent
Drop any one of these and it's something else — a lookup tool, a script, a chatbot. All three together are what the word actually describes.
It Reads State
It looks at real, current information — a record, a message, a document — instead of working from a fixed script written in advance with no awareness of what's actually in front of it.
It Decides, Inside Rules
It makes a judgement call about what to do next, but only within boundaries someone set on purpose — not by improvising freely outside them.
It Acts
It carries out the decision somewhere real — updates a record, sends a reply, files a document — rather than handing back a suggestion for a person to execute manually.
The Part That Makes It Trustworthy: Knowing When to Stop
A fourth behaviour matters as much as the first three, even though it's easy to leave out of a quick definition: a well-built agent recognises when a case is outside what it should decide alone, and hands it to a person instead of guessing. That boundary isn't a limitation bolted on afterward — it's part of what "within rules it was given" means from the start.
Definition Questions
What is an AI agent, in one sentence?
Software that reads information, decides what to do about it within rules it was given, and then actually does it in one of your systems — instead of just describing what should happen next.
Is "AI agent" just a rebrand of automation?
No, though the line can blur. A traditional automation follows a fixed path with no real judgement in the middle. An agent makes a judgement call within boundaries, which is what lets it handle a task with some natural variation rather than only the exact path someone scripted in advance.
Does an agent need to be conversational?
No. Conversation is one interface an agent can have, not a requirement. An agent reading incoming invoices and filing them never has a conversation with anyone, but it's still reading, deciding and acting — the three things that make it an agent.
Can an agent be wrong?
Yes, and any explanation that implies otherwise is selling something. The engineering work is in narrowing where mistakes can happen and making sure a wrong action is easy to catch and reverse, not in pretending the software is infallible.
Is every chatbot secretly an agent?
No — most aren't. A chatbot that only answers questions from a knowledge base and stops there isn't acting on anything. See AI Agent vs. Chatbot for the full distinction.
See What This Looks Like for Your Task
The definition is the easy part. Whether an agent is the right fix for what you actually do every day is worth a real conversation.