Illustration 3 of 3

Inside an AI Agent

On its own, an LLM can only talk. An agent gives that same model hands and legs — and keeps you in charge of what it's allowed to do.

Same request. Two very different worlds.

Give each one a real librarian's task and watch what it can — and can't — actually do.

“Email every community partner who hasn't replied in 30 days.”
A plain chat box words only
Email every community partner who hasn't replied in 30 days.
I can't open your partner list or your inbox myself — but here's how you'd do it:

1. Export your partners with their last-reply date.
2. Filter to anyone past 30 days.
3. Draft and send each a friendly note.

Want me to write the email template for you?
It can describe the work beautifully. It just can't do any of it.
vs
agent — community outreach has hands & legs
The same model — but wired to your files and your inbox, working in a loop until the job is done.
{{ l.marker }} {{ l.text }}
Permission needed: send 4 emails?

Same brain on both sides. The difference is everything around it: tools to act with, a loop to keep going, and a permission gate that hands the final say to you.

What turns a model into an agent?

Start with a model that can only think and write. Then give it tools — and watch what it can suddenly do.

The model
it can think and write
{{ powersLabel }}
Right now, it only has language, like an amazingly trained brain, but no body to take action in the world.
{{ p.text }}
Tap a tool to give it to the model
{{ equipNote }}

How it gets the whole job done: the loop.

A chat box runs this once — think, then speak. An agent goes around it, again and again, until the work is finished.

Think
Act
Observe
{{ loopPassLabel }} {{ loopNodeName }}

{{ loopText }}

A personal note — my opinion, not science

If you want real fluency with AI — and the leap in what you can create that comes with it — there's no way around learning some technology.

Not programming. Nothing like it. If you're comfortable in Microsoft Office — somewhere from capable to power-user — that's about the technical footing this work asks for. Agents are where that footing starts to pay off. I'll always tell you when I'm sharing an opinion like this one, so the conclusions stay yours.

Where would you place yourself today?
Avoid techGet byOffice power-userTinkerWrite code
{{ levelTitle }}
{{ levelMsg }}
That's the module.

You've seen how an LLM turns words into meaning — and how an agent turns meaning into work you can actually get done.

Back to The Science of AI