The Parts of an AI System

Sarah Jane Ross
AI for Beginners: Use Agents With Confidence Sarah Jane Ross 03:31

When you ask an AI assistant to do something, like turning a folder of receipts into a spreadsheet, it feels like you’re talking to one thing. Behind the scenes, several parts are working together, and knowing which part does what helps you work out why a result went wrong.

Here’s what happens when you ask an assistant on your computer to turn a folder of receipts into a spreadsheet:

Three columns for you, the app, and the model, with twelve numbered steps zigzagging between them from your request to the saved spreadsheet.
The Receipts Request, Traced Through the Parts

The app sits in the middle of every step. The model decides what should happen next, and the app carries it out, checking with you before it saves the spreadsheet.

The Parts

Whichever AI tool you use, the same handful of parts is working together:

  • The model does the thinking. It reads what it’s given and writes a reply, or asks for something to be done. GPT, Claude, Gemini, and Grok are all families of models.
  • The harness is the app around the model. It decides what the model gets to see, carries out the actions the model asks for, and checks with you when it should. The ChatGPT and Claude apps are harnesses, and you’ll learn more about them in Models and Harnesses.
  • Tools are the single actions a harness can take for the model, like reading a file, searching the web, opening a website, or sending an email. The model can ask for a tool, but only the harness can use one.
  • The agent is the model and the harness working together on a task over many steps. When your assistant goes off and works through your receipts, that’s an agent.
  • Skills are written instructions for a particular kind of job, a bit like recipes. They help the agent do that job the same way every time.
  • The computer it runs on is where all of this happens. Sometimes that’s your own computer, and sometimes it’s a computer in the cloud that belongs to the company. You’ll see why that matters in Where Is Your Agent Working?

In short, the model does the thinking, the harness carries out the actions, skills provide the instructions, and the work happens on a computer somewhere, which isn’t always yours.

Can you tell the parts apart?

Next: Models and Harnesses

00:00 Welcome to the course. It explains how AI chat apps and agents work, so you can get better results from them and use them safely. You don’t need any technical background.

00:10 The first lesson covers the parts of an AI system. When you ask an AI assistant to do something, it feels like you’re talking to one thing. Say you ask it to turn a folder of receipts into a spreadsheet.

00:22 Behind the scenes, several parts are working together on that request. Once you can tell them apart, it gets much easier to work out why a result went wrong.

00:31 Let’s follow that request through. On the left, that’s you. You ask, and you approve. In the middle is the app. It’s also called the harness, and it’s the part that actually does things.

00:44 And on the right is the model, which does the thinking.

00:48 You ask it to put your receipts in a spreadsheet. The app passes that on to the model, along with its own instructions. The model wants to know what’s in the folder.

00:58 So the app uses a tool to list the files.

01:02 Next, the model asks to open each receipt, and the app uses a tool to open them. Now the model has what it needs. It asks the app to make a spreadsheet with the totals.

01:13 But before anything gets saved, the app checks with you. You tell it to go ahead. Only then does the app use a tool to save the spreadsheet.

01:23 The model writes back that the spreadsheet is saved. The app shows you that reply, and the job is done.

01:31 If you look at the whole picture, the app is in the middle of every step. The model decides what should happen next. The app carries it out, and it checks with you before it saves anything.

01:42 Put the model and the app together, and that’s the agent. In this example, all of it runs on your own computer.

01:49 Now let’s name those parts. You’ll find the same handful in every AI tool you use.

01:56 First, the model. That’s the part that does the thinking. It reads what it’s given. Then it writes a reply, or it asks for something to be done. GPT is a family of models.

02:08 Claude is another, and Gemini and Grok are two more. Then there’s the harness. That’s the app wrapped around the model. It decides what the model gets to see.

02:18 It carries out the actions the model asks for. And it checks with you when it should. The ChatGPT app is a harness, and so is the Claude app. You’ll learn more about harnesses in a later lesson.

02:30 The single actions a harness can take for the model are called tools. Reading a file is a tool. So is searching the web. Opening a website is another, and so is sending an email.

02:41 The model can ask for a tool, but only the harness can use one. When the model and the harness work together on a task over many steps, that’s an agent. So when your assistant goes off and works through your receipts, that’s an agent at work.

02:56 Skills are written instructions for a particular kind of job. You can think of them a bit like recipes. They help the agent do that job the same way every time.

03:05 And finally, there’s the computer it all runs on. Sometimes that’s your own computer. Other times, it’s a computer in the cloud that belongs to the company.

03:14 You’ll see why that matters later in the course. So, in short, the model does the thinking. The harness carries out the actions. Skills provide the instructions.

03:25 And the work happens on a computer somewhere, which isn’t always yours.

Become a Member to join the conversation.