Harness engineering treats the AI agent as a system, not a chatbot. Guides are stable instructions (CLAUDE.md, AGENTS.md, rules) that enter every context. Sensors capture state (files read, commands run, tests passing) for the agent to decide the next step. Hooks execute deterministic actions at specific moments (pre-commit, pre-deploy, on-error). The result: the agent makes fewer mistakes, output is consistent across sessions, and human review focuses on architecture and security — not line-by-line diff. It's the foundation of the half-loop model.
GLOSSARY
Harness Engineering
What is Harness Engineering?
Approach that structures AI usage with guides, sensors and hooks versioned in the repository, instead of relying on ad-hoc prompts per developer.
Other concepts
Want to see this in your product?
Describe the idea. The pipeline maps the solution and you approve before anything runs.
Share your idea