Here is a strange thing about the word "agent." It appears in a science-fiction film, and everyone pictures a trench coat, red eyes, and a machine that has decided humanity is a problem to be solved. Then it appears in a software blog post, and everyone assumes something nearly as dramatic — a program that thinks for itself, sets its own goals, and operates without any human in the loop.
Neither picture is right. The real thing is much calmer, and once you see it, rather obvious.
This is article four in the AI in Plain English series. If you are new here, the series uses one image throughout: the Librarian behind a mail slot. The Librarian has read every book ever written and is extraordinarily good at continuing whatever is written on a note. But they sit behind a wall. You slip a note in. A note slides back. That is the whole relationship. And, crucially, the Librarian forgets you the instant the note disappears.
The Librarian's awkward limitation
Here is the problem we kept bumping into in the last few articles. Suppose you want help planning a weekend trip and you ask: "What is the weather forecast for Manchester on Saturday?"
The Librarian cannot answer that properly. They have not seen today's forecast. They have not seen any forecast from the past year. Their knowledge stopped at some point in the past, and even if it had not, they cannot step outside and check. They are behind a wall. They have a mail slot. That is it.
Or imagine you ask: "Add up the numbers in this spreadsheet and tell me the total." The Librarian can write words. They can reason about words. But there is no spreadsheet they can open. There is no calculator on their desk. The note comes in, a note goes out, and the spreadsheet stays exactly where it was — on your desk, not theirs.
One more: "Read the PDF I just attached and summarise it." There is no PDF behind that wall. There is only the slot.
This is not a criticism of the Librarian. They are genuinely brilliant at what they do. But what they do is a narrow thing: finish notes. Everything else — going to check something, opening a file, pressing a button in another program — is outside their job description entirely.
Enter the Runner
So you hire help. Not another genius. An ordinary helper whose job is to stand at the slot for you.
Call them the Runner — and "agent" is the technical word for the same idea. An agent is a program that wraps around the Librarian and handles the legwork.
Here is how the handoff works. You have a question. Instead of writing it on a note yourself and pushing it through the slot, you hand it to the Runner. The Runner writes the note, passes it through the slot, and reads what slides back. If the Librarian's reply says something like "I would need today's weather data to answer this," the Runner goes and fetches that data. Then the Runner writes a new note — your original question plus the weather data stapled to the back — and pushes it through again. This time the Librarian has everything they need, and the reply is useful.
The Runner then hands that reply back to you. From your point of view, the system "just answered."
This is the picture: you - Runner - Librarian. The Runner stands between you and the wall and handles everything the Librarian cannot.
The key line, stated plainly
Here is the insight that unlocks the whole concept.
An agent is all the parts that do not need a genius.
Read that again, because it removes the mystique immediately. The Librarian is the genius. They do the hard, subtle, creative work: understanding language, reasoning, generating coherent replies. An agent does everything else — the fixed, mechanical, ordinary steps. Look up a weather website. Open a file. Run a number through a calculator. Check whether a task is finished. Write results into a table.
None of those steps require brilliance. A school prefect could follow instructions to do them. The agent is, essentially, a very organised school prefect who carries messages back and forth and does small errands between trips.
How simple the early ones were
This is the part that surprises people. Given the dramatic language around AI agents, you might expect them to be enormously complex pieces of engineering.
Some early agents were almost laughably simple. In a basic setup, all the "agent" did was add a few extra sentences to the top of your note. Something like: "You are a helpful assistant. If you need to search the web, write SEARCH: followed by your query and I will bring you the results." That was it. The "intelligence" was just an instruction written in plain English. The agent was a loop: send note, check if the reply contained the word SEARCH, if yes go and search, then send again.
Later agents became more sophisticated — they could use many tools, make sequences of decisions, and break big tasks into smaller sub-tasks. But the principle never changed. The genius stays in the Librarian. The mechanical steps stay with the Runner.
Why the word "agent" sounds scarier than it is
Language matters here. The word "agent" carries a lot of baggage from films and novels where an artificial being develops its own desires and turns against its creators. That story is vivid and people remember it.
But in software, "agent" comes from a much older and more boring tradition. It just means "something that acts on behalf of something else." A travel agent acts on your behalf. A literary agent acts on behalf of an author. A software agent acts on behalf of a program — in this case, it acts on behalf of the Librarian, doing the errands they cannot do themselves.
When you hear "AI agent" in a news article, you can translate it directly: "a helper program that fetches things and presses buttons so the AI model does not have to."
There are genuine questions worth worrying about as these systems grow more capable — about oversight, about mistakes compounding when one agent hands off to another, about who is responsible when something goes wrong. Those are real conversations. But the concept itself, at its core, is a helper running errands. Nothing more.
The picture to hold onto
Imagine three desks in a row.
Your desk is on the left. The mail slot is in the middle of a wall. The Librarian is behind the wall on the right, out of sight.
Between you and the slot, the Runner sits at a small table. They have a phone for web searches. They have access to your files. They have a calculator. They have a list of instructions telling them what to do when the Librarian's reply contains certain phrases.
You hand a question to the Runner. The Runner handles everything else. You get an answer. The Librarian never left their chair. You never had to figure out how to use the phone or find the right file. The Runner did the non-genius parts.
That three-desk picture is the whole of what an AI agent is.
What comes next
The next article in this series — How AI reads files and searches the web — looks more closely at two specific errands the Runner can run: fetching a book from the right shelf by meaning (what engineers call RAG, retrieval-augmented generation) and going outside to ask around (web search). Both of those expand what the Librarian can usefully answer, but neither of them changes what the Librarian actually is. The genius stays behind the wall. The Runner does the legwork.
That is the pattern. The Runner is just a very organised helper with a short list of things they can do. The key line holds: an agent is all the parts that do not need a genius.
This is article 4 of 10 in the AI in Plain English series. Start from the beginning at aitutors.me or continue to the next piece above.