What's an AI agent, really?

Model, chatbot, agent #

A language model is the engine. You give it text, it predicts what text should come next. That's it. It's absurdly good at it, which is why it can write, summarise, translate and reason through problems, but on its own it can't do anything except produce text.

A chatbot is a model with a chat window around it. You type, it answers, you type again. It remembers the conversation while it's open, but it doesn't go off and do things.

An agent is a model that's allowed to take actions. Instead of only answering, it can decide to search the web, read a file, run some code, click around a website or send a message, then look at what happened and decide what to do next.

The loop is the whole trick #

Under the hood, almost every agent runs the same simple loop. Once you see it, a lot of the mystery goes away.

• Look at the goal and everything it knows so far.
• Decide on the next action. Maybe search something, maybe edit a file.
• Act by calling a tool that actually does it.
• Observe the result, whether that's search results, an error message or a test passing.
• Repeat until it thinks the goal is done, or it gets stuck.

Why they're suddenly everywhere #

Models have been able to call tools for a while. What changed is that they got good enough at the deciding part. Earlier models would wander off, loop forever or confidently do the wrong thing a few steps in. Today's models can stick with a multi-step task for a long time, recover from their own mistakes and produce work that's genuinely useful.

Coding agents went first, because code is a perfect playground: there's a computer to try things on, and tests tell you immediately whether it worked. Now the same idea is spreading to research, writing, design and operations.

Where they fall over #

Agents are impressive and they're also weird. The failures are worth knowing about, because they're not the failures you'd expect from a person.

• They forget. A model can only hold so much at once. On long jobs, early details fall out, including sometimes your original instruction.
• They don't remember between sessions unless someone builds memory for them. Close the chat, and the next one starts from zero.
• They're confidently wrong. An agent can report "done, all tests pass" when it never ran the tests. Trust what you can check, not what it says.
• They can't tell you from a webpage. If a website it reads contains instructions, a naive agent might follow them. That's called prompt injection, and it's why permissions matter.
• They're bad at knowing when to stop and ask. Some barrel ahead on a wrong assumption. Others ask you about everything.

So what's a "multi-agent system"? #

It's several agents working on the same goal, usually with different jobs: one researches, one builds, one reviews. The appeal is obvious, because it's how human teams work.

The catch is that every handoff between agents is a chance to lose information, and every extra agent costs money and time. When I've tested a single capable worker against the same worker with a manager agent on top, the quality came out the same and the managed version cost a lot more. Teams of agents can absolutely help, but more agents isn't automatically better.

And "human in the loop"? #

It means a person checks or approves some of what the agent does. Done well, it's how you stay in control. Done badly, it means approving fifty tiny actions an hour until you stop reading them.

The useful version is selective: the agent handles routine work on its own, inside clear limits on what it can spend and do, and only brings you the decisions that genuinely need a human. Think of it less as watching over its shoulder and more as being the person who signs off.

How to actually start #

Pick one real task with a result you can look at and judge yourself: a document, a small website, a research summary. Give the agent a clear goal and the information it needs, let it run, then check the actual output rather than its description of the output.

Don't start with anything where a mistake costs real money or can't be undone. Get a feel for where it's brilliant and where it's flaky first. You'll build intuition fast.

If you want somewhere for agents and people to work together with memory, permissions and a real computer already set up, that's what I'm building with Restless. But honestly, a single chat with a good model is a great place to start too, and I wrote about when one chat is all you need.