The Harness Design Philosophies of Claude Code and Codex

Subtitle: convergence, divergence, and the engineering choices behind two coding-agent systems.

“Life is suffering, and harness engineering is the art of making suffering deliberate.”

This book is a comparative anatomy of two harnesses: Claude Code, which grew out of operational experience, and Codex, which was born from a highly explicit policy and control layer. We are not tallying tools; we are examining where each system offloads trust, structures governance, and constrains a fundamentally unreliable model.

Three assumptions that guide this story:

• The comparison is centered on harness design, not model capabilities.
• A harness is a deliberate power-distribution system, not a feature checklist.
• Engineering differences usually emerge from where order is placed, not what it is called.

Recommended reading order:

1. Reading Map: How to read this book alongside the first volume
2. Preface: Two harnesses, not one accessory pack
3. Chapter 1: Why we pair Claude Code and Codex
4. Chapter 2: Two control planes—dynamic prompt assembly vs. structured instruction fragments
5. Chapter 3: Heartbeat vs. threads, rollouts, and state bridges
6. Chapter 4: Tools, sandboxes, and execution policy
7. Chapter 5: Skills, hooks, and local governance
8. Chapter 6: Delegation, verification, and persistent state
9. Chapter 7: Convergence, divergence, and where the philosophies overlap
10. Chapter 8: If you build your own harness, what to study first
11. Appendix A: Source map for this comparison
12. Appendix B: Checklists for spotting which harness you are closer to

For the short version of the thesis, jump to Chapter 7.