What is an AI agent?

A normal AI assistant answers questions. You ask, it replies, the conversation ends.

An agent is given a goal instead of a question. It works out the steps, does them, checks the result, and keeps going until it is finished or stuck.

That is the entire difference: answering versus doing.

Skip straight to checking what your assistant can reach

A concrete example

Ask a chat assistant “is my website down?” and it will explain how you might check.

Give an agent the same goal and it will actually check: fetch the page, look at what came back, and tell you. If the page is broken it might look at the logs. If you gave it permission, it might restart something.

Same model (the AI system that writes the replies) underneath. The difference is that the agent has tools and permission to use them.

The three things an agent needs

A model. The part that decides what to do next. This is the same kind of AI you have used in a chat window.

Tools. Ways to affect the world: reading a file, calling a website, running a command, sending a message. Without tools an agent is just a chatbot with ambition.

Memory. Somewhere to keep what it learned, so tomorrow’s run knows about today’s. Most of the frustration people have with agents comes down to poor memory.

Can it reach your computer?

This is the question people get stuck on, and it is separate from how clever the model is.

A chat window in a browser tab is sealed off from your machine by design. It cannot see which programs you have installed, it cannot open your files, and it cannot change a setting for you. Ask it to set something up and it will write you instructions, which is genuinely useful, but you are still the one doing the work.

An assistant that runs on your computer is a different thing: a coding agent in your terminal (a window where you type commands) or editor, or a desktop app you granted access to. It can look at what is actually there and act on it. That is not a smarter model, it is a different install with different permissions.

Two things sound like this but are not:

• A local model is about where the thinking happens, not what the thing can touch. See what is local AI. A local model answering inside a sealed chat window still cannot open a file.
• Web search or browsing in a chat tool reaches the internet, not your computer.

Check what yours can reach

You do not have to guess. Paste the prompt below into the assistant you are using right now, exactly as it is written, and let it work the answer out and report back. If it is a plain chat window, the honest admission that it cannot see your machine is the answer, not a failure.

Paste this, as it is. Nothing to fill in.

Please check whether you can actually access this computer, and tell me plainly what you find.
Do not change anything.

Run one harmless read-only test, such as showing me the name of the folder you are working in
or the name of my home folder. If you cannot run it, say so rather than guessing.

Then tell me in plain English:
- whether you can read files on this computer, and whether you can run commands
- what you cannot see or do
- if you have no access at all, which permission I would need to turn on or which tool I would
  need to install to get it

Work this out yourself. Do not ask me what kind of computer I have, and never ask me for
passwords, keys, codes or any other secret.

Reading the answer. If it shows you a real folder name, it can reach your machine, and the rest of its reply tells you how far that goes. If it explains how you could check yourself, or describes a folder without ever looking, it is sealed off from your computer and you would need something installed on the machine instead.

Where they genuinely help

Narrow, repetitive, well-defined jobs. Checking something on a schedule. Processing files that arrive in a folder. Doing the same five-step thing you do every Monday.

Where they still disappoint

Anything vague. “Grow my business” produces confident nonsense. Anything where being wrong is expensive, because agents fail quietly. They do not stop and say “I am not sure”, they carry on and produce something plausible.

And they are slower than you expect. Several steps means several model calls, and each one takes seconds.

Running one yourself

Hosted agents are the easiest way to see whether the idea suits you. If you want one running permanently, and you would rather your data stayed somewhere you control, you can run one on your own server.

That is the path the VPS guide walks through, from an empty server to an agent that stays online.

The one setting that matters

Give an agent the narrowest access that lets it do the job. Not your whole email account: one folder. Not your whole server: one directory. An agent that can only do a small amount of damage is an agent you can relax about, and relaxing about it is the entire point.