Free preview · no account needed
Welcome
You already use AI. You write prompts, run chains, maybe wire up a tool or two. This phase teaches the part underneath all of that: the harness.
One word is about to change meaning, and it's better said out loud. So far, an agent meant a written brief inside a chat tool. From here, agent means a coding agent: software that edits files, runs commands, and works in a repository. Same word, bigger machine. The first two lessons draw that picture before anything is built on it.
The harness is everything you build around a model — the context you feed it, the tools you let it call, the loop it runs in. Swap Claude for GPT, Gemini, or an open model and the harness barely changes. That's the point. The model is a commodity you rent. The harness is what you engineer, and it's the difference between an agent that's reliable and one that's flaky on the same task.
Four modules ahead — the first introduces the machine, the rest each engineer a different layer of the harness:
- The coding agent — what it is, the repository it works in, and a first supervised run you approve step by step.
- The mindset — treat the context window as a budget. Decide what the model sees, and what it never sees.
- The static layer — instruction files and skills the model loads before it starts. AGENTS.md and friends.
- The runtime — tools, loops, isolation, and verification. What the model can do, and how you check it.
Everything here is tool-neutral and portable. We name Claude Code, Codex, Cursor as examples, never as the only way. AGENTS.md is the default because it works across them. Learn the harness once and it carries to whatever model you run next year.