How AI Actually Works
Every day there’s AI news. A new model is out. Everyone’s talking about agents, prompts, sessions and context. You hear Claude and ChatGPT in the same breath as Opus 5.5 and GPT-6. A model, an agent, a prompt, a session: what do they all mean, and how are they different?
The model is the brain. Like a brain at birth, it starts out knowing nothing and has to be trained. A model is trained on a huge amount of text, guessing the next word over and over, and then taught to follow instructions. But unlike a person, who keeps learning for life, a model stops learning when training ends. You find out what it knows the way you would with a person: ask it something and judge the answer. Everything else in this piece, the apps, the tools and the agents, is built around it.
The app is the face, the model is the brain. ChatGPT and Claude are apps: the window you type into. Behind each one sits a model doing the thinking. ChatGPT runs on OpenAI’s GPT models. Claude runs on Anthropic’s models, such as Opus. Claude is also the family name for those models, which is half the confusion. The app decides what the model sees and what it can reach. The model decides what to say.
The prompt is what you say to it. Every question, instruction or file you hand over is a prompt. The brain can only work with what it’s given, so a vague prompt gets a vague answer. Think of briefing a smart new colleague who knows a lot about the world and nothing about your situation. A later piece covers what a good brief contains.
The session is one conversation. When you open a chat, you start a session. A new chat starts a new one.
The context window is how much it can hold. Everything the model sees at once has to fit in a space with a limit, called the context window. When it fills, most apps summarise the older part and carry on, and detail gets lost along the way.
Tools are its hands. On its own, a model can only talk. Tools let it reach the world: search the web, open a file, run a command, use another app. When a chat app looks something up for you, it’s using a tool. (Tools plug in through a standard called MCP, which a later piece covers.)
Get Daniel Ayia’s stories in your inbox
Join Medium for free to get updates from this writer.
An agent is a model with hands and a to-do list. Give the model a goal, a set of tools, and permission to act in a loop, and it stops just answering and starts doing: reading files, running commands, searching, checking its own work, and repeating until the goal is met. Same brain. Now it can act.
Subagents are helpers it delegates to. For a big job, an agent can spin up smaller agents, each handed one focused task and its own clean slate, then collect their results. It’s delegation, and it keeps the main conversation from filling up with side work.
“It learns from me.” Unlike a person, who keeps learning for life, a model stops learning when training ends. Nothing you tell it afterwards changes it. Your chats may help train the next model, unless you opt out in settings, but never the one you’re talking to.
“It remembers our conversation.” It re-reads it. Every time you send a message, the app hands the model the whole chat again, and it reads it from the top.
“It remembers me between chats.” Some apps keep notes on you and paste them into new chats. It feels like memory, but it’s the app passing notes. The model only ever sees what it’s shown.
“It lives on my laptop.” It runs on the company’s computers in large data centres. The app sends your words there over the internet and brings the answer back.
“It sounds sure, so it’s right.” Like a person, it can sound sure and still be wrong. Check what matters.
A model generates responses. An agent uses a model to take actions and work toward a goal.
Originally published at https://ayiaware.substack.com on November 5, 2026.
