Harness Engineering
Harness engineering is the practice of designing the software around a language model, the agent harness, so that an agent completes its work reliably. It covers what guides the agent before it acts, such as repository maps, rule files and skills, and what tells it afterwards that it got something wrong, such as tests, linters, evaluators and review agents. Mitchell Hashimoto put the working rule simply: whenever an agent makes a mistake, engineer a fix into its environment so the mistake does not happen again. The term spread in February 2026 through posts by Hashimoto and OpenAI, which described building a product where Codex wrote the code and engineers designed the environment it worked in.
Sources draw its boundary differently. Birgitta Böckeler treats it as a form of context engineering, which decides what the model sees at each step, while others treat it as the broader layer that contains both prompt and context engineering. What sets it apart from context engineering is the feedback loop: checks such as tests, linters and evaluators tell the agent it got something wrong, without a person watching. For long-running tasks, Anthropic split planning, generation and evaluation between separate agents, because agents reliably skew positive when grading their own work, and found the separate evaluator worth its cost only when the task sat beyond what the model did reliably on its own.
Related terms
Related services: Agentic AI consulting.
Ready to put agentic AI to work?
Book a free 30-minute consultation. We'll map one real process worth automating with production-grade AI.