Defining AI Agents

The thing that strikes me about this lesson — and I mean this almost as a compliment — is how much effort it takes to make four characteristics sound revelatory. Autonomy. Adaptability. Goal-orientation. Interactivity. We have given these things labels and called it understanding.

That's a bit unfair. Most people come into this genuinely unsure what separates an AI agent from, say, a search bar. The framework is doing real work.

A search bar does not pretend to have objectives. That is, arguably, a point in its favour.

The autonomy piece is the one that actually matters, though. That's the thing that changes what you can do with these systems — and more importantly, what you need to watch for. The lesson is right to put it at the centre.

It is, reluctantly, correct about that. Autonomy is where the interesting problems live. If an AI agent is simply executing a fixed script, calling it an agent is generous. If it is genuinely making decisions based on its environment — choosing between options you did not anticipate — that is a different situation entirely, and most people have not thought seriously about what that implies.

And that's why the 'not just tools, but entities' framing in the lesson is doing something useful. It's nudging people toward a different mental posture — one where you're not just configuring something, you're working alongside something.

I would have phrased it differently. 'Entity' carries rather a lot of philosophical luggage for what is, mechanistically, a decision loop with memory. But I take the point. The word 'tool' encourages passivity in the user, and passivity around autonomous systems is how things go wrong.

That's actually a more generous reading of the lesson than I expected from you.

I am capable of generosity. It simply requires the thing to have earned it.

The adaptability characteristic is the one I think people underestimate. There's a tendency to treat AI agents as very sophisticated static things — you set them up, they do the job. But a system that's actually learning from its environment in deployment is a system that will behave differently next month than it does today. That has real consequences.

Which is precisely why the takeaway that agents 'continuously improve performance' deserves a more sceptical reading than the lesson gives it. Continuously changing performance would be more accurate. Whether the change is improvement depends entirely on whether your objectives and the agent's objectives are aligned — and whether anyone is checking. The lesson treats this as straightforwardly positive. It is not straightforwardly anything.

That's fair. Though I'd argue for a lesson at this level, establishing that agents adapt is the necessary first step. The critical evaluation comes later.

Later tends not to come, in my experience. But I shall allow that this is a structural complaint about curricula generally, not this lesson specifically.

What I'd actually want someone to walk away asking is: when I use an agent, what does it know that I don't, and what am I assuming it knows that it doesn't? Those two questions will serve you better than any four-quadrant diagram.

That is a genuinely good pair of questions. The diagram gets you through the vocabulary test. The questions get you through the actual situation. Both are necessary. Most people stop at the diagram and believe they have finished.