This repository hosts two books on "Harness Engineering," addressing the critical challenge of integrating AI coding agents into real-world engineering systems. It targets engineers and researchers by providing a framework for designing robust control structures that ensure boundary, continuity, and consequence control, moving beyond basic prompt engineering to systemic AI governance. The benefit is enabling reliable, sustainable operation of AI agents in complex environments.

How It Works

Harness Engineering focuses on designing control structures that constrain and guide AI agents, treating prompt, tools, permissions, state, recovery, agent verification, and team processes as integral "organs" of a system. The core approach emphasizes that AI models are inherently unstable and prone to errors, necessitating a robust framework for managing their behavior and consequences. This contrasts with traditional prompt engineering by focusing on the systemic organization of execution rather than just input formulation.

Quick Start & Requirements

The repository provides online reading of the books via GitHub Pages at harness-books.agentway.dev, with direct PDF downloads available for book1-claude-code.pdf and book2-comparing.pdf. Local generation involves building two Honkit sites using npx --yes honkit build (requiring Node.js/npm) and then assembling them into a unified static site using a provided Python 3 script (tools/book-kit/build_pages_site.py).

Highlighted Details

• Explores "Harness Engineering" as a discipline for AI agent control, distinct from prompt engineering.
• Book 1 (Claude Code 设计指南) details runtime skeletons, control planes, query loops, tool permissions, context management, error recovery, and multi-agent verification.
• Book 2 (Claude Code 和 Codex 的 Harness 设计哲学) compares Claude Code and Codex, analyzing differences in their control plane designs, query loop responsibilities, and governance mechanisms.
• Introduces AgentWay (agentway.dev) as a practical extension for learning paths, exercises, and projects based on the book's concepts.

Maintenance & Community

The project points to agentway.dev for practical application and further learning, suggesting a connection to a broader initiative. No specific community channels (e.g., Discord, Slack) or details on core contributors/sponsorships are provided within the README content.

Licensing & Compatibility

No license information is specified in the provided README content.

Limitations & Caveats

The primary focus is on the necessity of robust control structures due to the inherent instability and error-proneness of AI models in engineering contexts. The text does not explicitly detail limitations of the repository's content itself, such as alpha status or specific unsupported platforms, but frames the problem space as one requiring significant engineering discipline.