LLMs vs AI Agents
Before we start using Copilot, it helps to understand what's actually going on under the hood. There are two technologies at play here: LLMs and AI agents. Let's look at both.
What's an LLM?
An LLM — Large Language Model — is the core technology powering pretty much all of the AI innovation we've seen over the last few years. The simplest way to think about it is as a next-word prediction machine. You give it some text, and it predicts what text should come next.
These models are able to do this because they've been trained on enormous amounts of text — we're talking the internet, books, documentation, all kinds of stuff. All of that reading allows them to predict, given some input text, what the most likely next bit of text would be.
Here's a quick example. Say we feed an LLM the phrase: "the cat sat on the". The model will return a few options for what could come next, each with a probability attached. Something like: mat (80%), roof (10%), chair (5%), floor (5%). Most of the time you'd go with the highest probability option — mat — but you can also tune the model to pick more creative, lower-probability responses when you want varied output.
Tokens & Vectors
Now, technically the model isn't predicting the next word — it's predicting the next token.
A token is just a smaller unit of text, which could be a whole word or just part of one. Under the hood, all of that text gets converted into numbers (called vectors) through a process called embedding, and the model stores all of this in a kind of high-dimensional geometric space.
The position of each token in that semantic space encodes meaning — which is how the model understands relationships between concepts.
But at a high level, just think of it like this: text in, text out. That's the LLM.
What's an AI Agent?
An AI agent is basically a software layer built on top of an LLM. The LLM is the brain — it does all the thinking — but the agent is what gives it hands. It's what lets the AI actually do things.
Where a raw LLM just takes text and returns text, an agent can:
Read and write files on your file system
Run terminal commands
Execute and reason about code
Run your tests
Do research on the web
Connect to external services and APIs
Orchestrate multi-step processes
So the agent is constantly feeding the right pieces of context into the LLM and using its responses to take real actions in the world — like writing code, editing files, or running a server.
There are a bunch of AI agent frameworks out there right now — GitHub Copilot, Claude Code, Cursor, Windsurf, and more. They're all built on this same concept, and they all work with models from companies like Anthropic, OpenAI, Google, and others. The concepts you learn here will transfer.
Next Steps
Now that we know what LLMs and AI agents are, let's start using Copilot. We'll look at how to pick a model and walk through all three modes — Ask, Plan, and Agent — by building a game from scratch.
Copilot Basics: Models & Modes