How an AI agent works: model, instructions, tools and memory

Forty messages before breakfast. Picture Lina. She runs a small online shop selling handmade candles, and she also books candle-making workshops. Every morning she opens her phone and finds about forty new messages. Some ask about shipping, some want to book a workshop, a few ask for a price on a wedding order, and one is an angry refund request. Answering all of them takes her two hours before she has made a single candle. What if a helper answered most of those messages overnight, booked the workshops, and left only the tricky ones for her? That helper is an AI agent. In this lesson we'll open one up and look at the parts inside, so you know exactly what you're building.

What you will learn. Here's the plan for the next few minutes. First, you'll learn the four parts that every AI agent is made of, whether you build it in n8n, Make, Zapier or with a few lines of code. Second, you'll see what each part actually does, using Lina's candle shop as the example every time. Third, we'll follow one real customer message, a workshop booking, through the full loop, step by step, and even run a tiny version of it. You don't need any background in machine learning for this. If you've ever trained a new employee, you already understand most of it.

The four parts of an agent. Every agent you'll build in this course has the same four parts. It all starts when a customer writes in. At the centre sits the model, the part that reads text and produces text. Feeding into it are the written instructions, which tell it who it is and how to behave. Then come its tools, the actions it's allowed to take, like checking a calendar or creating an invoice. And finally there's memory, what the agent knows about this customer and this conversation. Now watch one message travel through the agent. The question comes in, the model reads its rules, calls a tool to look something up, gets the result, saves a note, and sends a reply.

Part one: the model. The model is the brain of the agent. It's a large language model, or LLM, which simply means a program trained on huge amounts of text until it got very good at one thing: predicting which words come next. Think of the autocomplete on your phone, but strong enough to write a polite, accurate reply to a customer. Because it has read so much, it knows grammar, tone and plenty of general facts. But here's the important part: it knows nothing about Lina's shop. It doesn't know her prices, her workshop dates or her refund policy, and it can't open her calendar or send an email by itself. On its own, it can only read and write text.

Part two: the instructions. The instructions are how you turn a general model into Lina's assistant. They're usually called the system prompt, which just means the hidden text the model reads before every customer message. Think of it as the job description you'd hand a new hire on day one. It says who they are, for example the friendly assistant of a candle shop. It says what the goal is, like answering questions and booking workshops. It lists the rules, such as never promising refunds and passing angry customers to Lina. And it says how replies should look. Remember that the model reads this whole job description again with every single message, because it doesn't keep it from last time. We'll write these properly in the next lesson.

Part three: the tools. Tools are what let an agent do things instead of just talking. A tool is any action your workflow can perform: check the workshop calendar, book a slot, look up an order, create an invoice. Here's the part people often miss. The model never runs the tool itself. It writes a short request, something like "please run check calendar for Tuesday at ten". Your no-code workflow, say in n8n, sees that request, runs the real action, and passes the result back to the model. Only then does the model write the reply to the customer. This is what makes agents safe to sell: you decide exactly which tools exist, so the agent can only take the actions you wired up.

Part four: memory. The last part is memory, and it surprises most people. The model forgets everything the moment it finishes a reply. If a customer says "make it two people" in their second message, the model has no idea what "it" refers to, unless you send the first message again. So your workflow keeps the memory. Short-term memory is like a notebook on the front desk: the messages in this conversation, sent back in each time. Long-term memory is like a filing cabinet: a customer sheet or database with past orders and preferences, looked up when needed. Once you see that memory is just stored text you send back in, agents stop feeling like magic.

Worked example: one booking, in code. Let's make this concrete with a tiny version of the loop. You don't need to read this code line by line; just follow what each piece does. In n8n, these same steps would simply be boxes in a workflow. The calendar is a dictionary, which is just a lookup table of time slots and names: ten o'clock on Tuesday is free, two o'clock is taken by Ana. The memory is a list, an ordered list of messages, and it's our notebook. A function is a named set of steps, and the fake model function stands in for the real AI: when it sees the word book, it asks for the check slot tool instead of answering. Our code, playing the role of the workflow, runs that tool, books the free slot, and writes the answer. Both the customer's message and the agent's reply go into memory. In the output you can see the tool request the model made, the conversation saved in memory, and Tuesday at ten now marked for the new client.

The whole loop at Lina's shop. Let's step back and see the whole loop as a scene. On the left is the customer, relaxing at home and typing, can I book a workshop for two on Tuesday? In the middle is the agent, which really means the model plus its instructions, reading that message alongside the earlier chat pulled from memory. It decides it needs to check availability, so it asks for a tool. On the right is the calendar, the real system your workflow checks and updates. The answer flows back to the agent, which writes a friendly confirmation and saves a note for next time. Four parts, one smooth loop, and Lina is still asleep.

Common mistake: asking the model to just know. Here's the mistake almost every beginner makes. They connect a model to their inbox, write "you are a helpful booking assistant", and expect it to know which slots are free and what a wedding order costs. It doesn't, so it guesses. And it guesses confidently, in perfect sentences. A confident answer that's simply made up is called a hallucination, a bit like a new employee who'd rather invent an answer than admit they don't know. The customer then shows up for a workshop that's already full. The fix is always the same: if the agent needs a fact, either put that fact in its instructions or give it a tool to look the fact up.

Recap. Let's wrap up: four parts, three takeaways. First, the model is a powerful reader and writer, but it knows nothing about your business and can't act on its own. Second, the instructions and the tools work as a pair: instructions turn the model into your assistant with a role, a goal and rules, while tools are the actions your workflow runs when the model asks. Third, memory isn't inside the model at all. It's text you store, like conversation history and customer records, and send back in every time. Model, instructions, tools and memory: keep these four parts in mind and any agent, in any tool, becomes easy to read.

Your exercise. Now it's your turn, and you don't need any software for this one. Pick a small business you know well: a hair salon, a tutor, a photographer, maybe your own side project. On a single page, draw four boxes. In the model box, write which model you'd use, such as GPT or Claude, and one thing it must not be expected to know about this business. In the instructions box, write the agent's role and three rules it must follow. In the tools box, list three actions it would need, like book appointment or send invoice. In the memory box, write what it should remember about each customer. Keep that page, because in the next lesson we'll turn your instructions box into a real system prompt.