The harness: the loop, made real

What this lesson shows

A harness wraps the frozen model in a loop: it proposes an action, a tool runs, repeat.

The question it opens with

A chatbot just talks. But tools like Claude Code and Codex edit your files and run commands. What's the extra piece?

The walkthrough, in the lesson's own words

• A chatbot just talks. Tap the piece that makes it edit your files.
• Each move: WHO acts, then its result flows back into the notes.
• Your turn now. Scrub the beats, or cut the loop and watch it stall.
• That loop is the agent. The model never changed.
• Right. The model itself is frozen, same predictor as always. The loop around it runs tools and feeds the results back. Let's watch it run.
• Not the model. A frozen model of any size still just predicts text, it can't touch a file. The loop around it is what runs tools and feeds results back. Let's watch it run.
• Claude Code, Codex, and Cursor are harnesses.
• Every agent, expense sheets or code: decide → act → observe → repeat.
• You ask a coding agent to fix a bug. It reads files, edits code, runs the tests, then keeps going on its own until they pass. How does plain text-prediction do all that?
• The harness is the loop: the same frozen model proposes a tool call, the harness runs it, and the result flows back into the model's notes for the next turn. The model still only predicts text, so anything that touches your files is the loop acting, not the model reaching out on its own. Swap the code for an expense sheet and nothing changes: read it, propose the fix, apply it, re-check the totals.
• Frozen for now. The notes fill up once the loop runs.
• Loop cut: the model proposed an edit, but nothing runs it and nothing feeds back. Wire it back on.
• Every result flowed back in here. That feedback is the whole trick.
• Each tool result joins the notes, the model's next move reads it.

Key takeaway

The 'agent' is the harness loop, not a smarter model, the same predictor you met, wrapped in tools.

What you can do after this lesson

You can explain what an agent harness is, the tool loop around a frozen model, using Claude Code / Codex as examples.

Prefer reading first? This lesson has a companion explainer: AI coding agents · Agent vs chatbot

Where it leads: That whole loop runs inside one finite thing: the context window.

This is the written summary. The lesson itself is interactive: you predict, drag and operate the mechanism above, and the reveal answers you.