wquguru/harness-books

Harness engineering in two books: constraints, query loops, context governance and multi-agent verification in Claude Code and Codex. harness-books.agentway.dev

What it is

Harness Books is a project that publishes two free online books about "harness engineering" — the practice of designing the systems that surround AI coding agents (like Claude Code or Codex) so they behave safely and reliably in real software projects.

The books are available to read online or download as PDFs, and the project also includes the source files and build tools to generate the sites locally.

What the books cover

  • Book 1: Harness Engineering: A Design Guide to Claude Code — uses Claude Code as a concrete example to explain the core components of a harness: prompts as control logic, the query loop, tool permissions, context management, error recovery, multi-agent verification, and team rules. It argues that the real risk of AI coding tools isn't occasional wrong answers, but a lack of structure to handle consequences.

  • Book 2: The Harness Design Philosophies of Claude Code and Codex — compares how Claude Code and Codex each place "order" in their systems. One relies more on runtime discipline, the other on a structured control layer. The book helps readers understand architectural trade-offs, how to evaluate agent systems, and what to study if building your own harness.

Core claims

  • Harness engineering is about how constraints organize execution, not just prompt engineering.
  • Once a code-writing model enters a real engineering environment, the main problem shifts from answer quality to behavioral consequences.
  • Prompts, tools, permissions, state, recovery, verification, and institutions are not add-ons — they are organs of a single control structure.
  • When comparing agent systems, the key question is where order is actually placed, not the feature checklist.
  • If a team can't turn individual experience into reusable rules, it will struggle to turn an agent into a stable system.

Who it's for

  • Developers, team leads, and architects who use or plan to use AI coding agents in real projects.
  • People interested in the architectural design of agent systems, not just usage tips.
  • Those who want to understand the differences between Claude Code and Codex at a deeper level.

How to read it

  • For the full framework: read Book 1, then Book 2.
  • If you already know coding-agent tools and want the architectural comparison directly: start with Book 2.
  • If you only want the conclusions: read Book 1 Chapter 9 and Book 2 Chapter 7.

Additional resources

The project also points to AgentWay, a separate practice platform where the ideas from the books are extended into training paths, drills, and project exercises.

Technical details

The repository includes the book source files (in Markdown), build scripts (Python-based, using Honkit), and instructions to build the full site locally. The final output is written to a dist/ folder. Both English and Chinese versions are available.

In short

Harness Books is a set of two practical, architecture-focused guides for anyone who wants to understand how to safely integrate AI coding agents into real engineering workflows — not just how to use them, but how to design the systems around them.

Related

  • Project
    cloudflare/skills

    A plug‑in library that adds Cloudflare‑specific “Agent Skills” to AI coding assistants (Claude Code, Codex, Cursor, VS Code, etc.). Each skill teaches the agent how to work with a Cloudflare product—Workers, Wrangler, Durable Objects, Turnstile, Cloudflare One, and more—by providing up‑to‑date docs, CLI commands, and API calls via Cloudflare’s remote MCP server. Install via the native plugin marketplace of the target agent or manually copy the skill folders.

  • Project
    mitsuhiko/agent-stuff

    A collection of reusable skills, extensions, and prompt commands for the Pi Coding Agent to enhance its capabilities and workflow automation.

  • Project
    vuejs-ai/skills

    A community‑driven set of AI‑agent “skills” that teach coding assistants how to write better Vue 3 code, covering best practices, router, Pinia, testing, debugging, and composable patterns.

  • Project
    leookun/cursor-byok

    cursor-byok is a local open-source gateway that lets Cursor use your own model APIs instead of built-in options, preserving agent features like tool calling and MCP.