An AI agent is an AI system that is given a goal and works towards it on its own, by taking actions, looking at what happened and deciding what to do next. A chatbot answers you once and waits. An agent keeps going, step after step, until it thinks the job is done. That loop of acting and checking is the whole idea, and it is also why agents go wrong in the ways they do.
What does an AI agent actually do?
Say we ask an agent to book a table for four at an Italian restaurant near the office on Friday. A chatbot would suggest a few places. An agent would search for restaurants, open their booking pages, check which have a free table, pick one, fill in the form and send us the confirmation.
Each of those steps follows the same pattern:
- Look at the current situation (the search results, the booking page).
- Decide on the next action, using a language model such as GPT, Claude or Gemini.
- Act, using a tool: a web browser, an email account, a spreadsheet, a calendar.
- Check the result against the goal, and go back to step 1.
The language model is the part that reads and decides. The tools are what let it do things in the world. The loop is what turns the two into an agent.
How is an agent different from a chatbot?
The difference is who closes the loop. With a chatbot, we do. We read the answer, judge whether it is any good and type the next message. With an agent, the system reads its own results and chooses its own next step. We only see the beginning and the end.
This is why agents have become the big AI story of 2026. In October, Microsoft launched a version of Copilot with agents that carry out work in the background, and Google began testing a work agent that can have its own Workspace account, email address and spending limit. With these, we hand over a task and come back later for the result.
Where does the idea come from?
The idea is older than computers. In 1943, three researchers, Arturo Rosenblueth, Norbert Wiener and Julian Bigelow, wrote a short paper called Behavior, Purpose and Teleology. They argued that a machine could behave with purpose if it compared where it was with where it wanted to be, and used the difference to correct its next move. They called this negative feedback.
A thermostat works this way. It measures the room, compares it with the setting and switches the heating on or off. Wiener had spent the war on anti-aircraft guns that did the same with the path of a plane. The paper became one of the founding texts of cybernetics, the study of goal-seeking systems in machines, animals and societies.
An AI agent is the same loop with a language model in the middle. The goal is written in words, the measurements are web pages and files, and the corrections are actions. So the questions cybernetics asked about thermostats and gun sights apply to agents as well, and they explain most of what goes wrong.
Why do AI agents go wrong?
The checking is weak. A thermostat measures the room with a thermometer. Many agents check their work by asking the same model whether the work looks right. If the model misread the booking page the first time, it can misread it again when it checks. A loop is only as good as its measurement, and here the measurement comes from the same place as the mistake.
Small errors add up. Every step depends on the one before. If an agent gets each step right nine times out of ten, a ten-step task will go wrong more often than it goes right. Long tasks are where agents are most useful and where they fail most.
They get stuck in loops. An agent that cannot tell it is failing may try the same action over and over: reloading a page, rewriting the same file, sending the same search. This is what a feedback loop does when the signal that should stop it never arrives.
The goal is fuzzy. “Book a nice restaurant” leaves the agent to decide what nice means. A thermostat has a number to aim for. An agent often has a sentence, and it fills the gaps with guesses.
The world has more surprises than the agent can handle. The cybernetician W. Ross Ashby put this as a rule in 1956, the law of requisite variety: a controller can only keep a system on track if it can respond to as many different situations as the system can throw at it. A pop-up, a changed form or a fully booked evening are all variety the agent has to match. When it cannot, it either stops or carries on with the wrong plan.
How do we keep agents under control?
Mostly by adding feedback from outside the agent. Spending limits, like the ones on Google’s new agent, stop a loop from running up costs. Asking a person to approve key steps, such as paying or sending an email, puts a human back into the loop where it matters. Giving the agent fewer tools and a narrower task reduces the variety it has to deal with. Each of these is an old cybernetic remedy: a better measurement, a limit on the loop, or a smaller problem.
Common questions
Is ChatGPT an AI agent? The standard ChatGPT chat is a chatbot: it answers and waits. ChatGPT also has agent features that can browse the web and carry out tasks on their own. The same model can work either way, depending on whether it is allowed to act and loop.
What is agentic AI? “Agentic AI” is the general term for AI systems that act on their own towards a goal. An AI agent is one such system.
Are AI agents safe to use? They are safest on tasks where a mistake is cheap and easy to undo, and where we can check the result. For anything involving money, private data or messages to other people, it is wise to keep a human approval step.
Will AI agents replace jobs? Agents can already take over some routine, step-by-step office tasks. How far that goes depends a lot on how reliable their checking becomes, which is the weak point described above. We look at this question in more detail in a separate piece.
What is the difference between an AI agent and a bot? A traditional bot follows fixed rules written by a programmer. An AI agent uses a language model to decide its next step, so it can handle tasks nobody wrote rules for, and it can also make mistakes nobody predicted.
For more on the loop at the heart of all this, see our guide to feedback and the toolkit question What is the feedback structure?