devpilot-harness-engineeringlisted

# Harness Engineering ## Overview An **agent harness** is everything around a coding agent *except the model itself*: the guides that steer it before it acts, the sensors that catch drift after it acts, the context it reads, the sub-agents it delegates to, and the automation that maintains code quality in the background. **Core principle:** When agents write most of the code, the engineering team's primary job shifts from *writing code* to *building the harness that lets agents write good code*. Constraints you'd normally postpone until hundreds of engineers are onboard become day-one prerequisites. **Operating definition (Mitchell Hashimoto):** every time the agent makes a mistake, engineer the harness so it cannot make that mistake again. The harness is never "done" — it grows from observed failures, not speculation. **Synthesized from** OpenAI's "Harness engineering: leveraging Codex in an agent-first world" (a ~5-month experiment by a team growing from 3 to 7 engineers, producing roughly 1M lines of code and ~1,500 merged PRs with no manually-written code), Martin Fowler's harness-engineering article, and HumanLayer's "Skill Issue" writeup. ## When to Use **Use when:** - Starting a repo where agents will author most code - Agent output quality is decaying (style drift, inconsistent patterns, duplicated utilities) - You're spending more time reviewing agent PRs than the agents spent writing them - Deciding where to invest: AGENTS.md, skills, sub-agents, linters, fit