"AI agent" is the phrase of the year, and almost nobody hands you a clean definition before using it. You see it in product launches, LinkedIn posts, and breathless headlines about software that "works while you sleep" — and most of us nod along, quietly unsure whether an agent is a genuinely new thing or just ChatGPT wearing a smarter hat.
If that's where you are, you're in good company. I spent the back half of last year assuming "agent" was marketing gloss for "chatbot, but we'd like more money for it." It isn't. There's a real, specific difference underneath the hype, and once it clicked for me, a lot of the noise suddenly sorted itself into "this matters" and "this is a demo that will never survive contact with real work."
This piece is part of our Terminology Tamer series — the same plain-English treatment we gave large language models. By the end, you'll be able to define an AI agent in one sentence, picture what it's actually doing, and tell the difference between a tool worth your attention and a shiny one worth ignoring.
The one-sentence answer
An AI agent is a system that takes a goal you give it, then works out the steps needed to reach that goal and carries them out — using tools, checking its own progress, and looping until the job is done.
That's the whole idea. The key word is does.
A chatbot talks. You ask, it answers, the exchange ends, and nothing has changed in the world except that there's now some text on your screen. An agent acts. You give it an outcome — "find me three suppliers, compare their prices, and draft an email to the cheapest" — and it goes off and works through that, taking actions on your behalf rather than just describing what you could do.
Keep that distinction nearby: a chatbot produces words, an agent produces results. Most of the confusion around this topic comes from quietly sliding back into thinking of an agent as "a chatbot that's a bit cleverer." It isn't more clever. It's wired up differently.
How to actually picture what an agent does
Here's the mental model I wish someone had given me first.
Think about the difference between asking a knowledgeable colleague a question versus handing a task to a capable intern.
When you ask the colleague a question, you get an answer. "How would I put together a competitor pricing comparison?" — and they talk you through it. Useful. But you still have to go and do the work.
When you hand the task to the intern, something different happens. You say, "Can you pull together a pricing comparison for our top three competitors by Friday?" They don't just describe the process back to you. They go away. They open a browser. They look things up. They notice one competitor's pricing page has changed and adjust. They put it in a document. They come back when it's done — or when they're genuinely stuck.
That second pattern is what an AI agent is reaching for. Underneath, it's still powered by the same kind of language model that runs a chatbot. The difference is everything wrapped around the model: the ability to use tools, to take a series of actions, and to keep going without you prompting each individual step.
And this is the part that changes how it feels to use one. With a chatbot, you're in the driving seat the whole time — prompt, read, prompt, read. With an agent, you hand over the wheel for a stretch and check the work when it returns.
🧠 Quick Challenge: True or false — an AI agent is just a chatbot that gives longer, more detailed answers.
- A) True
- B) False
Answer: B) False. The difference isn't the length or detail of the reply — it's that an agent takes actions (uses tools, works through steps, loops until done) while a chatbot only produces text for you to read. As we covered above: a chatbot produces words, an agent produces results.
The loop that makes an agent an agent
If you want the one technical idea that separates a real agent from a chatbot, it's this: an agent runs in a loop.
Roughly, the cycle looks like this:
- Take the goal. You give it an outcome to aim for, not a single instruction.
- Make a plan. It breaks the goal into smaller steps.
- Take an action. It does one step — searches the web, reads a file, sends a request to another piece of software, writes some code.
- Check the result. It looks at what came back. Did that step work? Is it closer to the goal?
- Work out what's next. Carry on, adjust the plan, or stop because the job's done.
Then it goes back to step three and does it again. And again. Plan, act, check, adjust — until it reaches the goal or hits a wall it can't get past.
That loop is the whole trick. A chatbot does one pass: text in, text out, done. An agent cycles through that loop as many times as the task needs, working out what to do next at each turn based on what just happened.
The two things that make the loop possible are worth naming, because they're what you're really paying for when you pay for an "agent":
- Tools. The model is connected to things it can actually use — a web browser, your calendar, a spreadsheet, a database, another app's controls. Without tools, a model can only generate text. With them, it can take action in the real world.
- A degree of autonomy. It's allowed to take the next step without stopping to ask you each time. How much rope it gets is a setting, and a genuinely important one — we'll come back to that.
Where you'll actually meet agents
This is where it gets less abstract. You may already be using something agent-shaped without anyone calling it that. A few examples a working professional is likely to bump into:
- A research assistant that takes a question, searches multiple sources, reads what it finds, and comes back with a synthesised summary and links — rather than one best-guess answer from memory.
- A coding assistant that doesn't just suggest a snippet but reads your project, makes changes across several files, runs the code to see if it works, reads the error, and fixes it — looping until the thing runs.
- A "do this for me" workflow built into a tool you already use: "Book the cheapest flight that lands before noon," or "Go through these 200 customer emails and tag the ones that mention a refund."
- A scheduling or inbox helper that doesn't just draft a reply but checks your calendar, proposes times, and sends the invite once you say yes.
Notice the shared thread: in each case the model is taking actions in other systems, not just producing a block of text for you to copy and paste. That's the line.