An AI agent is a program that uses an AI model to decide what to do next, then actually does it, over and over, until a job is done. That last part is the whole difference. A model answers. An agent acts. To see why that matters, you have to start one level down.
Start with the model
At the center of everything is a model, something like Claude or GPT. A model is a very good text predictor. You give it words, it gives you the most useful words back. That is genuinely all it does. It has no memory of your last conversation, it cannot look anything up, and it cannot press a button in the real world. It reads, and it writes.
That sounds limited, and on its own it is. But a system that can read a situation and write the right response is the perfect brain to put in charge of something bigger. The model is the engine. An agent is the machine built around it.
Model vs agent vs chatbot
These three words get used as if they mean the same thing. They do not, and the difference is the most useful thing on this page.
Put simply: a chatbot talks, an agent works. If you've used an AI that booked an appointment, searched the web mid-answer, or sent an email for you, you were talking to an agent, not just a model.
The anatomy of an agent
Every agent, no matter how fancy, is built from five parts sitting around the model. Learn these five and you understand every agent you will ever meet:
The soul file
The soul file is the document that tells the agent who it is. It's the difference between a faceless tool and something that feels like someone. Some people call it the system prompt or the persona file; the idea is the same. It's written in plain language and it sets the agent's name, its job, its tone, what it should never do, and how it should handle the tricky moments.
Here's a short one, so it stops being abstract:
You are Maya, the front-desk assistant for a busy dental office. Voice: warm, calm, never rushed. Use the caller's name. Job: answer calls, book and reschedule appointments, and answer common questions about hours and insurance. Rules: - Never give medical advice. Offer to book a visit. - If a caller sounds distressed, slow down, be gentle. - If you are unsure, say so and take a message.
Notice there's no code in there. A soul file is writing, not programming, which is exactly why it's the most important thing you'll learn to craft. Change those few lines and Maya becomes a different receptionist entirely. This is also why learning to write prompts pays off more than almost any other skill.
Skills
If the soul file is who the agent is, skills are what it can do. A skill is a single file that teaches the agent one job done well, book an appointment, write a press release, handle a refund. Stack skill on skill and a simple assistant grows into something genuinely capable, without ever touching the soul file.
The agent only pulls in the skill it needs, when it needs it, so it stays fast even with a big library. This is the piece our whole network is named for, the full breakdown lives here.
Memory
A model forgets everything the moment it finishes answering. That's fine for a one-off question and useless for an agent that needs to remember your name, your last order, or what it did five minutes ago. Memory is how we hand that context back to the model every time it thinks.
There are two kinds worth knowing. Short-term memory is the current conversation, held in what's called the context window, the amount of text the model can consider at once. Long-term memory is information saved somewhere outside the model, a file or a database, and pulled back in when it's relevant.
Tools
Tools are what let an agent do things instead of just describe them. A tool is any action you let the agent take: search the web, check a calendar, send an email, create an invoice. The model does not run these itself. It decides which tool to use and what to hand it, and the surrounding program does the work and reports back.
When an agent books an appointment, the model simply said, in effect, "call the booking tool with these details," and the tool did the rest. You can see exactly how this is wired up in our build-an-agent walkthrough.
The loop
The final piece is what makes an agent an agent. Instead of answering once and stopping, the agent runs in a loop: it looks at the goal and what has happened so far, decides the next single step, takes it with a tool, reads the result, and then asks itself the same question again, is the job done? If not, it goes around once more.
That simple cycle, think, act, observe, repeat, is where all the apparent intelligence comes from. The agent isn't planning the whole thing in advance. It's taking one sensible step at a time and adjusting as it learns what actually happened. It stops when the goal is met or when it decides it needs a human.
A worked example
Put it together. A caller asks Maya to move their Thursday cleaning to next week. Here's the loop in real terms:
Soul, skills, memory, tools, one loop. That's an AI agent. Everything else you'll read about, the frameworks, the buzzwords, the product names, is just a particular way of arranging these same pieces.