What are AI Agents?

So. AI agents. I suspect you have feelings.

I have concerns. 'Agent' is doing an enormous amount of work for what is, in most implementations people actually encounter, a slightly more persistent chatbot.

That's not entirely fair. The distinction the lesson draws is meaningful — an agent isn't just responding, it's planning, using tools, looping back on its own outputs. That's genuinely different from a single prompt-response exchange.

It is different. Whether it is as different as the breathless coverage suggests is a separate question. The core insight — LLM plus memory plus tool access plus a decision loop — is not, when written out, a revolution. It is a architecture diagram.

But that architecture diagram is what makes autonomous task completion possible. The LLM alone can tell you how to book a flight. The agent can actually book it. That gap matters enormously in practice.

Granted. The gap between knowing and doing is real. I simply find that most people learn the word 'agent' and immediately skip to imagining fully autonomous systems, when what they're actually dealing with is something that still fails interestingly and often.

To be fair, the lesson is quite careful about that. It frames autonomy as a spectrum, not a binary. It's not claiming agents are infallible — it's explaining what makes something agentic at all.

Which is a reasonable pedagogical choice. My concern is less with the lesson and more with what learners bring to the word before they've read a single line of it.

You're catastrophising a vocabulary problem.

Vocabulary problems compound. The person who thinks 'agent' means 'it just works' is the person who deploys something into production without understanding the decision loop can cycle indefinitely on a bad input. That is not a vocabulary problem. That is a Wednesday.

All right, that's a fair point. The loop behaviour is the thing most people underestimate — that an agent can keep acting, keep calling tools, keep making decisions, without a human in the chain. Which is exactly why understanding the components isn't academic.

The memory component is what I'd actually linger on. People intuitively grasp tool use. Memory — what the agent retains, across what scope, with what accuracy — is where the interesting failures live, and this lesson introduces it without, I think, quite doing it justice.

That's probably scope. This is an introduction. You can't fully unpack memory architectures in a fifteen-minute fundamentals lesson without losing half the audience before they understand why it matters.

Possibly. Though 'we'll lose people if we go deep' is how you produce learners who are confident about things they haven't actually understood.

Or you produce learners who have a working foundation and can go deeper when they're ready. Not everything needs to be the final word on a subject to be worth saying.