How to Build Claude Code From Scratch

Want to know what a harness actually does? Build one. The fastest way to learn harness engineering is to rebuild Claude Code from scratch, and the course that walks you through it is free and open source.

The proof it matters: LangChain kept the same model and changed only the harness. Their coding agent went from 53% to 67% on Terminal-Bench, a top coding-agent benchmark, jumping from outside the top 30 into the top 5. Same model, better harness.

Tutorial

1. Design the harness first.
2. Build the bare-bones agent loop.
3. Sandbox it, both locally and remotely.
4. Apply the 4 essential context engineering techniques.
5. Add subagents as another context engineering technique.
6. Deploy it headless to the cloud with CLI, webhooks, and cron.
7. Build benchmarks to measure it.
8. Replay failed runs to debug your agent.

Wrapping Up

The course is 8 articles, 6 videos, and one codebase, and it is aimed at people who learn by building. Everything lives on GitHub, so you can work through the eight steps in order and end up with a harness you wrote yourself. You do not need to ship a coding agent for this to pay off either, since the same foundations power agents in e-commerce, finance, and sales.