What is an AI agent?

An AI agent is an AI that is given a goal instead of a question, and works out for itself how to reach it. It gathers context, does something, checks its own result and goes round again, until the task is done or it hits something it cannot decide on its own.

That is the whole definition. The rest of this guide is about that one loop, because the loop is what separates an agent from an answer. Almost every AI tool you use today runs a version of it.

There is no clean divide between chat and agent any more. There is a spectrum of how agentic the tool is.

The loop

Think of a capable colleague given a task, who does not come back to you for every paragraph. The agent reads what exists, drafts, looks it over, fixes what does not hold and delivers when it is right. If it asks you something along the way, it is because something was unclear, not because it needs a yes at every step.

It is three steps and an arrow back.

Gather context. The goal and what it has been given. A sentence from you, a folder of files, or both.

Act. It reads, writes, searches, runs a tool or makes a file.

Check. It holds the result up against the goal. Do the numbers add up, is everything there, is it actually what you asked for.

Improve. That is the arrow back. If it does not hold, the agent adjusts and goes round again. Or it stops and asks, because something could not be decided.

Say the task is a client summary from a folder of meeting notes. The agent reads, writes a page, checks it against the notes and comes back with the page and one question about a number it cannot place. That is one round of the loop.

All AI is agentic now

ChatGPT searches the web, Claude reads the files you upload, and Copilot writes and runs Python in Excel to work on a spreadsheet. All of them take several steps and use tools before they answer, which is the same loop.

Andrew Ng, who has taught a whole generation machine learning, proposed in The Batch in June 2024 that we stop asking whether something is an agent and ask how agentic it is instead. It is a degree, not a switch. I find that the most useful way to look at it.

So the difference is not whether the loop is there. It is how many rounds it runs without you, and what it is allowed to touch while it runs.

The spectrum

The tools can be lined up by how much they do on their own.

Chat that answers. You ask, it answers, and the loop runs through you every round. That is still the right choice for a quick question or a sentence to rewrite.

Chat with tools. It searches, reads a file or calculates before it answers, but you are there for each step. This is where most people are today, often without noticing they are using an agent.

An agent with a workspace. You point it at a folder, describe the result and let it work for a long time. Claude Cowork, Claude Code and Codex are here, and so is Copilot Cowork, which works on files in the cloud.

An agent that runs without you. The same agent runs on a schedule and reaches your calendar, inbox or systems through connectors. A morning brief can be waiting before you are at your desk, and you are on the loop, reading the result.

Several agents. One agent splits the task and hands the pieces to other agents that work in parallel.

The same tool can sit in several places on the scale. Claude is chat with tools when you ask a quick question, an agent with a workspace when you point it at a folder, and an agent that runs without you when you set up a scheduled task. The scale is about the setup, not about the brand name.

Agents working in parallel

In Claude Code and Codex, one agent can start several others, let each run its own loop in its own context, and assemble what they report back.

You come in where it matters, on the goal at the start and on the check where something is unclear or expensive to get wrong. More agents means more tokens and more to read, so it pays off only when the task genuinely comes apart into pieces.

How to build a team of AI agents shows how to set one up in Claude Code.

Where you stand

The question that comes up at every bootcamp is where the person should stand while the agent works. It matters more than how agentic the tool is.

In the loop you sit at the check step. Every round waits for your yes. The agent drafts, you read, you say go on or change this, and only then does it act again. For a new task that is the place to stand. You learn what it gets right and what it gets wrong, and it learns from the corrections you save.

On the loop you set the goal and the boundaries, say which results you want to see, and let the agent take the rounds. Afterwards you read what came out, and the flags it left where it was unsure.

The price of standing in the loop is that the loop runs at your speed. If you have approved eight drafts in a row without changing a word, and the ninth is waiting while you are in a meeting, you are ready to move out. No one moves all the way out at once. What works is releasing one task at a time.

New bottlenecks

The work does not disappear when the agent takes the rounds, it moves. Making things went from expensive to cheap, while deciding what to make, and checking what comes back, costs what it always did. Peter Drucker drew the line between doing things right and doing the right things. The agent has become very good at the first.

Three new bottlenecks show up. The brief, because an unclear task gives you twelve agents moving quickly in the wrong direction. The check, because more comes back than before and somebody has to read it and check the quality. And access, because an agent without context, files and permissions stops at the first question.

Three starting points

Something has to set the loop going, and in practice it is one of three things.

You ask. You open the tool, describe the task and stay while it works, which is where everyone begins.

A pulse. The agent wakes on a schedule, looks for work and does what it finds, and if there is nothing to do, it does nothing.

An event. Something happens, a file lands in the folder or an email arrives, and the agent starts by itself.

The last two are the ones that move you out of the loop. You no longer start the work, you read the result of it. Both need the agent to run somewhere that is awake when you are not. More on that in Cowork moved to the cloud.

What an agent needs

Four things around the agent matter more than the model it runs on.

A goal it can check itself against. Say what you want to end up with and what should happen when something is unclear, not how it should work.

A workspace. A folder with the sources, the examples and the rules, so the context sits there instead of being explained again every time. We call it an agent in a folder.

Access. Connectors to the calendar, the inbox and the systems you use, so it can fetch what it needs itself. More on that in connectors and automation in Cowork.

Your way of working. A skill that says how you do this particular task, so it does not start from generic assumptions. More on that in skills in Claude Cowork.

Tasks that fit

Not everything deserves an agent. For a simple one-off question, chat is faster. For rule-based work that happens often, ordinary automation is cheaper and safer. An agent gives you the most where the task repeats and takes judgement, because that is where a company holds a lot of knowledge that was never written down.

Reading invoices into a spreadsheet and turning meeting notes into a project brief both fit, because the material is ready and the work can be described.

An example

Fifty invoices sit in a folder as PDFs. They need to go into a spreadsheet with supplier, date, amount and account, and you want to know which ones the agent was unsure about.

What you write.

YouCopy

Read every invoice in this folder and make invoices.xlsx with one row per invoice, columns supplier, invoice date, amount excluding VAT and account. The accounts are in accounts.md. If you are unsure about a row, include it anyway and write what is unclear in the unsure column.

The first thing the agent does is make its own task list.

Gather context. The agent opens the folder, sees fifty files, reads the account list and the first few invoices to find the pattern.

Act. It goes through invoice by invoice and writes the rows.

Check. It counts the rows against the number of files, sees that two invoices are scans with no text layer and that one is in another currency.

Improve. It takes another pass at the two scans, gets what it can, and marks the currency row instead of converting on its own.

What comes back is the spreadsheet, three flagged rows and a couple of sentences about what it was unsure of. You read three rows, not fifty. It is the same loop from the top of this guide, with your own files in it.

Try it tomorrow

Pick one task where your own files are the point. Put them in a folder. Open Claude Desktop, give Claude access to the folder, and write what you want to end up with, where it should be saved and what it should do when something is unclear. On Pro and Max, Cowork is becoming part of the ordinary Claude conversation, so you may not see a separate Cowork option.

YouCopy

Read all the meeting notes in this folder and write a project brief summarising status, decisions taken and what is outstanding. Save it as project-brief.md here, one page at most. If two notes disagree, flag it and let me make the call.

Watch the first few rounds as they run, and notice where the agent stops and what it asks about. Read the file that comes out, and fix the brief rather than the file.

If you would rather learn it with other people, we run Claude Code bootcamps in Oslo, where we work hands on with agents all day.

Questions we get

Is ChatGPT an AI agent? Yes, and how agentic it is depends on how you use it. A chatbot answers what you ask, while an agent acts, checks its own result and goes round again. In ordinary chat the loop runs through you every round, which puts ChatGPT at chat with tools. ChatGPT Work, its agent for longer tasks, takes several steps on its own across your files, connected apps and the web. Codex, from the same company, is an agent with a workspace that works in the codebase.

What is the difference between an AI agent and RPA or ordinary automation? Automation follows rules you have written and does the same thing every time. An agent is given a goal, picks the steps itself and can stop and ask. RPA and other automation fail quietly when the world stops looking like the rules. Use automation where the rules are fixed, and an agent where the task takes judgement.

What does agentic AI mean? That the AI runs the loop itself, in several steps, with tools, towards a goal. The word covers everything from chat with web search to agents that work for hours without you, so it says little on its own.

Do I need to know how to code to use an AI agent? No. Ask Claude or ChatGPT to analyse a spreadsheet and it can write a small program, run it and give you the answer, without you reading the code. Claude Cowork and tools like it are built for anyone with files and tasks, and you need code only when you set out to build software yourself.