OpenAI의 하네스 엔지니어링 방법론을 어떤 저장소에 설정합니다. 어떤 AI 코딩 에이전트와 함께 작동 - Copilot, Codex, Claude, Cursor, 또는 다른 사람.

그것은 무엇을

대리인 첫번째 기술설계 환경에 어떤 repo를 변형하십시오:

• AGENTS.md — ~100라인 프로그레시브 디플로이저 맵(encyclopedia)
• Architecture layer - 기계적으로 의존성 경계
• **테스트 인프라 ** - 테스트 러너, 유틸리티, 아키텍처 경계 테스트
• **CI 파이프라인 ** - lint, typecheck, 테스트, 모든 PR에 빌드
• **Golden 원칙 ** - 코덱에 대한 canonical DO/DON’T 패턴
• Garbage collection — 안티-패턴 편류를 위한 주간 자동화된 검사
• Pre-commit Hooks - CI에 도달하기 전에 위반

지원된 더미

어떤 더미도 함께 작동합니다. 기술 자동 탐지 및 적응:

설치

harness-engineering

An agent skill that sets up OpenAI’s harness engineering methodology in any repository. Works with any AI coding agent — Copilot, Codex, Claude, Cursor, or others.

What It Does

Transforms any repo into an agent-first engineering environment:

• AGENTS.md — ~100-line progressive disclosure map (not an encyclopedia)
• Architecture layers — mechanically enforced dependency boundaries
• Testing infrastructure — test runner, utilities, architecture boundary tests
• CI pipeline — lint, typecheck, test, build on every PR
• Golden principles — canonical DO/DON’T patterns for the codebase
• Garbage collection — weekly automated scans for anti-pattern drift
• Pre-commit hooks — catch violations before they reach CI

Supported Stacks

Works with any stack. The skill auto-detects and adapts:

Install

npx skills add charlesanim/harness-engineering

Or to install globally (user-level, available in all repos):

npx skills add charlesanim/harness-engineering -g

Usage

Once installed, just ask your agent:

“Set up harness engineering in this repo”

Or any of these triggers:

• “Make the repo agent-ready”
• “Add architectural boundaries”
• “Set up testing and CI”
• “Adopt agent-first engineering”

What is Harness Engineering?

Coined by OpenAI after shipping ~1 million lines of agent-generated code in 5 months with zero manually-written code. The core insight: don’t write better prompts — design better environments.

8 Principles:

1. Engineers become environment designers
2. Give agents a map, not an encyclopedia
3. If agents can’t see it, it doesn’t exist
4. Enforce architecture mechanically, not through instructions
5. Boring technology wins
6. Entropy management is garbage collection
7. Throughput changes merge philosophy
8. Agent-to-agent code review

Read the original: openai.com/index/harness-engineering

License

MIT