What it does

This repo publishes two free books on “harness engineering”—the architecture around code-writing AI agents. Book 1 dissects Claude Code’s runtime: query loops, tool permissions, context governance, failure recovery, multi-agent verification. Book 2 compares Claude Code against Codex, mapping where each system places order (prompt assembly vs. instruction fragments, runtime discipline vs. structured control layers). Both are readable online or as PDFs; the repo also contains build tooling to assemble the Honkit-based sites.

The interesting bit

The core argument is that prompts aren’t chat-window sugar—they’re part of the control plane. The real risk isn’t an occasional wrong answer; it’s a system with no structure for handling consequences. That’s a shift from “make the model smarter” to “make the system accountable.”

Key highlights

• Book 1 traces why agent systems need components like control planes, recovery paths, and team-level rules—not just better prompts
• Book 2 compares Claude Code and Codex as divergent harness designs, not feature checklists
• Includes source maps grounding claims in actual files (Appendix C in Book 1, Appendix A in Book 2)
• Executable checklists for turning principles into constraints (both books)
• Built with locale-aware Honkit; local build scripts provided

Caveats

• The repo is primarily content and build tooling, not a runnable harness you can install
• No quantitative benchmarks or empirical evaluations—claims are architectural and conceptual
• The “Python” language tag is misleading; the build scripts are Python, the books are prose

Verdict

Worth your time if you’re building or adopting coding agents and need a framework for thinking about control, not just output quality. Skip if you want copy-paste prompts or a ready-to-deploy system.