What is a harness?

Read this short agent session and label each line C (context), T (tool), P (permission) or L (loop):

1. You: "Add a title to index.html"
2. Harness shows the model your message + the contents of index.html
3. Model: "I want to edit index.html"
4. Harness: "Allow edit to index.html?  [yes / no]"
5. You: yes
6. Harness edits the file and tells the model "edit done"
7. Model reads the result and decides it is finished

Write your labels on paper or in a note.

You're done when every line 2–7 has a letter and you can say in one sentence what the harness did that the model could not. (Suggested: 2 C · 3 T · 4–5 P · 6 T · 7 L. Some lines fit two letters, and that's fine.)

• The model is the brain. It only reads and writes text.
• The harness is the body around it. It does four jobs: context, tools, permissions, loop.
• OpenCode, Claude Code, Codex and Grok's CLI are all harnesses. We use OpenCode.

Next lesson (1.4): the context job up close. Why the model can only "see" so much at once.