The short answer
An agent is a language model that has been given tools and permission to use them in a loop: it decides what to do, does it, looks at the result, and goes again until it thinks the job is done. That is genuinely different from a chatbot. It is also far less reliable than the word “agent” suggests, because every step can go wrong and the mistakes compound.
A chatbot answers. An agent acts.
Ask a chatbot to book a table and it writes you a lovely paragraph about booking a table. It has no hands. Everything it produces is text handed straight back to you.
An agent is the same model with three additions: a set of tools it is allowed to use, permission to decide which to use, and a loop. It might search the web, open a page, fill a form, run a calculation, or call another program — then look at what came back and decide the next move.
That loop is the whole idea. Think, act, look at the result, think again. It repeats until the job is done, until it gives up, or until it hits whatever limit you set.
ORBIT / THE NAVIGATOR
The word covers an enormous range. Some “agents” are a model with a search box bolted on. Others really do run for twenty minutes and change things in the world. Always ask which one you are being sold.
What the loop looks like
Suppose you ask an agent to find the cheapest train from Leeds to London next Tuesday morning and put it in your calendar. A rough version of what happens:
One agent loop, step by step
- 1. Think
- “I need train times. I should search a rail site.”
- 2. Act
- Opens a rail site and searches Leeds to London, Tuesday morning.
- 3. Look
- Reads the results. Sees six trains and their prices.
- 4. Think
- “The 07:03 is cheapest. Now I need the calendar.”
- 5. Act
- Creates a calendar entry for the 07:03.
- 6. Look
- Confirms the entry was created, then reports back.
An illustration of the shape, not a transcript. Real loops are messier and often take many more steps.
Where it goes wrong
Every step in that loop is a place to be confidently wrong, and the errors multiply rather than cancel out. If each step is right nine times in ten, a ten-step task finishes correctly about a third of the time.
The failures also look different from a chatbot's. A chatbot invents a fact and you read it. An agent invents a fact and then acts on it — books the wrong day, emails the wrong person, deletes the wrong file. It is wrong at speed and out of sight.
And it cannot tell that it is stuck. A loop that is failing looks, from the inside, exactly like a loop that is working: it keeps deciding, acting and looking. This is why sensible agents have hard limits on how many steps they may take and how much they may spend.
BEEP / THE MYTH-BUSTER
My favourite failure is the confident loop. Twenty minutes of purposeful activity, each step following sensibly from the last, all of it building on something it made up in step two.
What they are genuinely good at today
Bounded, repetitive jobs with a checkable result. Reading fifty documents and pulling the same three fields out of each. Trying an approach, seeing the error, and trying again — which is why coding assistants were the first agents that properly earned their keep.
The common thread is that you can check the output quickly. If verifying the work takes longer than doing it yourself, the agent has cost you time while feeling like it saved some.
What to actually do
- Ask what tools it can reach, and what it is allowed to change.
- Give it a spending limit and a step limit before you give it a task.
- Start with jobs where a wrong answer is obvious and cheap.
- Keep anything irreversible — sending, paying, deleting — behind your own confirmation.
The questions worth asking before you let one loose
What can it actually reach? An agent with a calendar is a different proposition from one with your email, your files and a saved card.
What does it do when it is unsure? The good ones stop and ask. The rest guess and carry on, because carrying on is what the loop does.
And who is accountable when it gets it wrong? Not the agent. That answer has not changed and is not going to.
PIXEL / THE COACH
Give it the boring job you would happily check afterwards. That is the one where it earns its place and you learn what it is like before anything matters.
See it for yourself
Reading about it only gets you so far. This one is written to show you the thing described above, rather than describe it again.
Before you do anything else: list the tools you can actually use right now, what each one is allowed to change, and which of those changes I would not be able to undo.
Then paste it into whichever assistant you use — ChatGPT, Claude, Gemini or Copilot. Not sure which? Meet them all.
If you remember nothing else
- An agent is a model with tools and a loop: think, act, look, repeat.
- Errors compound across steps, so long tasks fail far more often than single answers do.
- It is wrong at speed and out of sight, which is worse than being wrong on a page.
- Best on bounded, repeatable work you can check quickly.
- Keep anything you cannot undo behind your own confirmation.
Where this came from
- Anthropic, Building effective agentschecked 26 September 2026
- OpenAI, A practical guide to building agentschecked 26 September 2026
Anything here that turns out to be wrong is worth telling us about. Figures that move — prices, sizes, energy use — are kept in the registry with their dates, rather than written into the prose where they would quietly go stale.
Questions about “AI agents”Asked by readers, answered by us. Open to read them.
Open to load them.
Ask about this page and we will answer it here, where the next person can read it too.
Nobody else can answer you, and you cannot answer anybody else. This is not a forum — we would rather say that than run a room we cannot watch at three in the morning.
Your question goes up with the answer and without your name. Nothing identifies you, ever.
If we cannot answer it we will tell you why, and that one stays between us.