Part I: Why, and What's Different

Chapter 1: Introduction — Why Move to Codex Now

Written: 2026-04-28 Last updated: 2026-04-28

1.1 The 24-Hour Gap

Here is the statistic that sets the context: in March 2026, Claude Code authored approximately 4% of all public GitHub commits — single-handedly [DEV Community, 2026]. In The Pragmatic Engineer's survey of 906 engineers, 46% named Claude Code as their "most loved" AI coding tool — the highest in the survey.

Twenty-four hours after GPT-5.5 launched on April 23, 2026, post #46 "From Claude Code to Codex" appeared [Um, 2026]. This book starts from that 24-hour gap.

The answer to "why move to Codex?" is not simply "GPT-5.5 is better." The more precise answer: GPT-5.5 launched inside Codex before it was available via API [Willison, 2026]. OpenAI's newest frontier model shipped into its harness product first. The harness beat the API to market.

When Simon Willison routed through a Codex subscription to access GPT-5.5 on launch day, that wasn't a workaround — it was confirmation that the unit of model deployment had shifted. The harness had stopped being a wrapper around the model. It had become the primary delivery channel.

Figure 1.1: Claude Code dominance, GPT-5.5 lands inside Codex, and the 24-hour gap to the migration post — the moment the harness beat the API to market. illustration by author Gemini assisted
Figure 1.1: Claude Code dominance, GPT-5.5 lands inside Codex, and the 24-hour gap to the migration post — the moment the harness beat the API to market. illustration by author Gemini assisted

1.2 The Harness IS the Product

This book has one core argument: the harness is the product.

"Harness" may be unfamiliar. Chapters 2 and 4 cover it in depth. For now: the harness is the code and configuration that determines what context the model sees, what tools it can use, and how it stores and retrieves knowledge.

Claude Code externalizes the harness. You write CLAUDE.md, configure subagents, attach hooks, create skills. The harness lives in your hands.

Codex internalizes the harness. AGENTS.md and ~/.codex/config.toml are the interfaces, but the harness logic is baked into the product.

OpenAI's decision to ship GPT-5.5 into Codex first is the most direct expression of this philosophy. Instead of releasing the model via API first, they embedded it in the harness product. The frontier model and the harness became inseparable.

1.3 The Paradigm Shift

Andrej Karpathy called the emergence of AI coding agents a "phase change" [Karpathy and Patel, 2025] — like water at 100°C suddenly becoming steam, the nature of coding work flips at some point.

The Claude→Codex move is the next step in that phase change. It is not a tool swap. It is a shift in how you think:

  • Before: "What should I tell Claude to do this turn?" — issuing instructions externally, turn by turn
  • After: "Write the rules in AGENTS.md; let Codex follow them across turns" — externalize the rules, internalize the autonomy

The real cost of this transition is not code changes. It is knowledge externalization.

Post #46 states it plainly [Um, 2026]: the Claude→Codex switch is possible when your project's rules, state, and next actions are in plain-text files — AGENTS.md, HANDOFF.md, TASKS.md — rather than locked inside conversation history or tool-specific memory. If those files exist, the switch is mechanical. If they don't — if knowledge lives only inside Claude conversations — the switch means starting over.

Figure 1.2: Knowledge externalization in practice — only plain-text artifacts (AGENTS.md, HANDOFF.md, TASKS.md) survive the Claude to Codex migration; tool-specific memory and conversation-bound context must be rebuilt. illustration by author Gemini assisted
Figure 1.2: Knowledge externalization in practice — only plain-text artifacts (AGENTS.md, HANDOFF.md, TASKS.md) survive the Claude to Codex migration; tool-specific memory and conversation-bound context must be rebuilt. illustration by author Gemini assisted

1.4 Who This Book Is For

Two assumptions underpin this book:

  1. You have used Claude Code. You have typed claude in a terminal. You know what CLAUDE.md is.
  2. You have heard of harness engineering but haven't applied it seriously. The word "harness" is familiar but your workflow doesn't yet depend on it.

Part II of this book is therefore not a migration guide. It is a fresh learning guide. You'll build a harness in Codex from the ground up — not porting Claude Code patterns 1:1, but learning Codex's way of doing things.

This assumption is calibrated to reality. By early 2026, "harness engineering" had entered Korean developer vocabulary — there were 40-minute crash courses on it [Korean Dev Blog, 2026]. But actually building a production harness workflow remained uncommon.

1.5 Map of the Book

Four Parts:

Part I: Why, and What's Different (Ch 1–3)

Why Codex, how the two tools differ, and your first Codex session.

Part II: Learning Harness Engineering Anew in Codex (Ch 4–6)

What a harness is, how to build one in Codex, and how to do multi-agent verified coding.

Part III: Practical Codex Guide in the GPT-5.5 Era (Ch 7–9)

What three model releases in 67 days changed, community patterns from the first month, and advanced multi-agent work.

Part IV: From LLM Wiki to the AI Scientist (Ch 10–12)

Karpathy's LLM Wiki pattern, your own activity as AI external memory, and AI Scientist workflows by domain.

Figure 1.3: Map of the book — 4 Parts and 12 chapters. Part I introduces; Part II is fresh harness-engineering learning in Codex; Part III covers the GPT-5.5 era; Part IV moves from LLM Wiki to the AI Scientist. illustration by author Gemini assisted
Figure 1.3: Map of the book — 4 Parts and 12 chapters. Part I introduces; Part II is fresh harness-engineering learning in Codex; Part III covers the GPT-5.5 era; Part IV moves from LLM Wiki to the AI Scientist. illustration by author Gemini assisted

Any chapter can be read independently. If this is your first pass, read Part I → Part II in order. If you've never run Codex, jump straight to Chapter 3's setup tutorial.

1.6 The "Before" Snapshot

Karpathy predicted in 2025 that AI agents would reach expert-programmer level in software development [Karpathy, 2025]. By the Jan-2026 newsletter snapshot [Bizzotto, 2026], that prediction was materializing.

In February 2026, Anthropic shipped Claude Sonnet 4.6 with SWE-bench Verified at 79.6% (80.2% with prompt modification) [Anthropic, 2026]. Developers reported: "better instruction following, less over-engineering." A paradox emerged: as models improved, the harness became more important, not less. The better the model, the more skill was required to frame the work correctly.

Then came April 23. GPT-5.5 shipped inside Codex.

This book records what came after.

References

  1. Dev.to, "Claude Code authored ~4% of all public GitHub commits in March 2026," 2026-03. [DEV Community, 2026]
  2. terryum, "From Claude Code to Codex," terryum-ai posts, 2026-04-24. [Um, 2026]
  3. Willison, Simon, "GPT-5.5," simonwillison.net, 2026-04-23. [Willison, 2026]
  4. Karpathy, Andrej, Dwarkesh Patel Podcast — AI coding agents discussion, 2025. [Karpathy and Patel, 2025]
  5. Karpathy, Andrej, "Expert programmers and the future of software," 2025. [Karpathy, 2025]
  6. Code with Andrea, "Jan-2026 Newsletter," 2026-01. [Bizzotto, 2026]
  7. Korean Developer Blog, "하네스 엔지니어링 40분 정복," 2026. [Korean Developer Blog, 2026]
  8. Anthropic, "Introducing Claude Sonnet 4.6," 2026-02-17. [Anthropic, 2026]
  9. Korean Developer Blog, "하네스 엔지니어링 40분 정복," 2026. [Korean Dev Blog, 2026]