Something confuses almost everybody at the start. People say “Claude Code is better than
ChatGPT at coding”, as if they are different brains.

Often they are using a very similar model. What differs is the harness: the program
wrapped around the model that decides what it can see and what it can do.

Three common harnesses:

Chat. You type, it replies with text. You are the hands. You copy the code, you paste
it into your editor, you save the file, you run it. The model never touches your computer.

Autocomplete. It suggests the next few lines while you type in your editor. It sees a
small amount of nearby code. Fast, narrow, no conversation.

Agent. The model is given tools: read a file, write a file, run a command, search
the project. Now it can act. You say “the tests are failing, fix them”, and it reads the
test file, edits the code, runs the tests, sees the result, and tries again.

The idea to take away

The model is roughly the same in all three. The difference in power comes from the harness
around it.

An agent is not smarter than a chat window. It just has hands. It can look things up
instead of guessing, which means it hallucinates less, because now it can check.

That last point is important. An agent that reads your actual file does not have to imagine
your file.

What this costs you

Hands cut both ways. A chat window that is wrong wastes your time. An agent that is wrong
can delete a file, install something you did not want, or push broken code.

That is why Module 3 spends a whole lesson on permissions before you run anything serious.

Try this before the next lesson

1. Ask a chat tool: “What is in the file package.json in my project?” Watch it fail
   honestly, or fail dishonestly.
2. Write down the last three things you asked an AI to do. For each one, mark whether it
   needed hands or only words.
3. Which of the three harnesses have you already used without knowing its name?