The librarian and the assistant

Ask a librarian for a book about volcanoes and she will tell you exactly where it is. Aisle 7, third shelf, blue spine. You still walk there and fetch it.

Ask a personal assistant the same thing and she comes back with the book in her hands.

The librarian gave you information. The assistant gave you a result. That is the difference between a chatbot and an AI agent.

Words in, words out

In The Helpers You Can’t See, a girl called Ava asks a chatbot to book a pizza table for two at seven. It replies with the restaurant’s address, its opening hours and its phone number. Everything it says is helpful, and the table is still unbooked.

A chatbot lives inside the conversation. It can explain, draft and summarize, but it cannot open a website or fill in a form.

Tools give it hands

An agent can reach outside the chat through tools: anything outside the model that it is allowed to use, such as a web search, a calendar, an email account or a booking form.

The model itself never touches any of them. It writes a request, like an order slip, in an agreed format. Ordinary software around the model, called the harness, reads the slip, runs the errand and pastes the result back in. Hello, AI compares it to a head chef who never leaves the kitchen and sends runners to the market.

A loop keeps it going

What makes an agent is a loop. It decides a step, uses a tool, reads what came back, and decides the next step with that new knowledge. Round it goes until the goal is met, a limit is reached, or a person steps in.

Anthropic, the maker of Claude, describes an agent as a language model using tools in a loop. The blogger Simon Willison put it this way in 2025: “An LLM agent runs tools in a loop to achieve a goal.”

There is no special agent brain. The same kind of model that drafts your emails sits inside, with tools, a loop and permissions wrapped around it.

Where you already meet them

Help desks use agents at scale. In early 2026 ServiceNow said agents inside the company resolve about 90% of its routine IT tickets, like expired passwords, without a person. Klarna said its AI assistant handled about two-thirds of its customer service chats in its first month in 2024, though the company later brought some human staff back.

Coding is where agents went furthest first. A coding agent can read the code, make a change, run the tests, see what failed and try again.

The research group METR measures how long a job agents can manage, in human working time. In early 2025 the best agents finished tasks that take a person about an hour half the time, and that length had doubled roughly every seven months from 2019 to early 2025. METR’s January 2026 update found it has doubled faster since 2023, about every four months.

One question helps you sort what you see. Who chooses the next step? If a developer fixed the steps in advance, it is a workflow. If the model chooses as it goes, it is an agent.

In one sentence: an agent is an ordinary language model given tools to act with and a loop that lets it keep going until the job is done.