What is harness engineering?
Harness engineering means improving the instructions, tools, skills, state, and checks around a coding agent. Agents equal model plus harness, and each mistake shows what the harness is missing.
The main idea
If you use Claude Code or Codex, you already use a harness. The same goes for Cursor, Copilot, Antigravity, and other agents. Once you use these agents for real work, you see that the model still makes mistakes and does odd things.
The blog by Birgitta Böckeler uses a simple equation. Agents equal model plus harness. The model reasons. The harness is everything around the model that helps it work and catches its mistakes. Harness engineering is the work of improving the basic harness your coding tool already gives you. You do not have to write code all day to do it. The harness should match the work you are doing, whether that is building an app or managing cloud systems.
In the video I show what a harness can contain: rules in agents.md or claude.md, skills, MCP tools, and state files with logs about the project. When the model makes a mistake, I ask what was missing from the harness, then I fix the environment so the next run goes better.
Key points
- Agents equal model plus harness. The model reasons, and the harness helps it work and catches mistakes.
- A harness can include rules in agents.md or claude.md, skills, MCP tools, and state files.
- OpenAI first used one giant agents.md file, then moved to a short map that points agents to the right document.
- Guides steer the agent before it acts. Sensors check its work afterward, such as a test or a cloud plan.
- If the agent forgets a decision, write it into a file the agent reads. If it says the app works while the page is broken, add a browser check. If a bug returns, add a test or a linter rule.
Save my free seat →