Agent Harness Skill

Language: English | 中文

Agent Harness is a platform-neutral AI collaboration skill. It distills practical harness engineering ideas into an installable protocol built around Planner, Generator, Evaluator, and Coordinator roles.

核心判断：

同一个模型、同一句需求，最终质量往往不只取决于模型本身，而取决于你在模型周围搭建了什么系统。

Core thesis:

With the same model and request, output quality often depends less on raw model capability and more on the system wrapped around it.

What Is Inside

• skills/agent-harness/SKILL.md: canonical portable skill for Claude Code, Codex, and other agent runtimes.
• skills/agent-harness/SKILL.zh-CN.md: Chinese reading version.
• skills/agent-harness/references/bootstrap-checklist.md: coordinator startup checklist.
• skills/agent-harness/references/artifact-templates.md: starter templates for the handoff artifacts.
• skills/agent-harness/scripts/harness-gate.sh: lightweight artifact gate for checkpoint/final validation.
• skills/agent-harness/scripts/harness-runner.sh: minimal external loop that checks the gate, writes continuation prompts, and can call a configured agent CLI.
• skills/agent-harness/scripts/harness-monitor.sh: self-started watchdog for the active Claude Code/Codex session; it writes a continuation prompt when the workspace goes idle before the gate passes.
• README.zh-CN.md: full Chinese explanation.
• README.en.md: full English explanation.
• docs/ARTICLE.zh-CN.md: Chinese public article draft.
• docs/ARTICLE.en.md: English public article draft.
• assets/cover-agent-harness.png: cover image for articles and sharing.
• MAINTENANCE.md: long-term maintenance workflow.
• scripts/sync-local.sh: sync the canonical skill into local Claude/Codex skill directories.

Core Pattern

Agent Harness is not "ask several agents to all write answers."

It is a single-agent system of role separation:

• Planner expands the user's request into a restrained high-level spec.
• Generator builds only against an agreed contract.
• Evaluator independently verifies quality and catches failures.
• Coordinator orchestrates handoffs, constraints, disputes, iteration, and final synthesis.

For codebase and product repair tasks, the default outcome is a safe first implementation slice unless the user explicitly asks for proposal-only work. Plans are scaffolding; verified product changes are the deliverable.

The method emphasizes:

• restrained upstream planning
• generation/evaluation separation
• iteration contracts
• calibrated evaluators
• file-based handoffs
• context reset through structured transition artifacts
• component audit and ablation
• simplified final review mode when stronger models make per-cycle structure unnecessary

Quick Install

Claude Code:

mkdir -p ~/.claude/skills
cp -R skills/agent-harness ~/.claude/skills/agent-harness

Codex:

mkdir -p ~/.codex/skills
cp -R skills/agent-harness ~/.codex/skills/agent-harness

Other runtimes:

Use SKILL.md as the canonical method document and map the role prompts and artifacts to the tools your runtime supports.

Local Maintenance

This repository is meant to be the canonical long-term project. Update the skill here first, then sync local installs:

scripts/sync-local.sh

Private source material and temporary review artifacts should live under private/, which is ignored by git.

Relationship to Harness Engineering

This repository is strongly informed by harness engineering practice, especially the ideas of restrained planning, strict external evaluation, contract-driven iteration, evaluator calibration, context reset, and component-by-component simplification.

It is not a transcript or a copied implementation. It is a rewritten, platform-neutral protocol meant to be installed, adapted, and audited across assistant runtimes.