From Claude Code to Codex
From external harness to internalized autonomy, and onward to LLM-Wiki-driven research
How Claude Code users move to Codex for multi-agent coding and AI Scientist workflows. — 4 Parts, 12 Chapters
First published: 2026-04-28 | Last updated: 2026-05-05
Two Tools, Compared
Seven-dimensional comparison: Claude Code (external harness) vs Codex (internalized autonomy).
Learn Harness Anew
Not migration — fresh learning. Set up multi-agent verified coding in Codex from scratch.
Updated Through 2026-05-04
Covers Codex CLI 0.128.0: /goal, permission profiles, import flow, and Worktree/Cloud operation.
LLM Wiki to AI Scientist
Karpathy's Obsidian workflow, markdown as external memory, domain-specific research automation.
Part I: Why, and What's Different
Introduction — Why Evaluate Codex Now
Why evaluate Codex without breaking Claude Code repositories: community views, knowledge externalization, and safe order
→ 02Two Tools, Compared — Philosophy and Interface of Claude Code and Codex
Memory, tools, autonomy, edit model, permissions, context, collaboration — side-by-side across 7 dimensions
→ 03Codex First Steps — Quick Adaptation Guide for Claude Code Users
Installation, ~/.codex/config.toml, first AGENTS.md, safe takeover of Claude Code repositories, and command mapping
→Part II: Learning Harness Engineering Anew in Codex
Harness Engineering at a Glance — Concepts and Core Patterns
Why a harness is needed, recap of core patterns from the Claude era (CLAUDE.md/subagents/hooks/skills)
→ 05Building a Harness in Codex — AGENTS.md, Config, Directory Structure
How to write AGENTS.md, configure ~/.codex/config.toml, preserve Claude Code harnesses during import/rebuild, and stand up the first multi-agent setup
→ 06Verifiable Vibe Coding — Splitting Work, Verification, and Review across Agents
Don't let one agent do everything — multi-agent gates for work, test, and review
→Part III: Practical Codex Guide in the GPT-5.5 Era
Model Release Timeline — Sonnet 4.6 → Opus 4.7 → GPT-5.5
Per-release changes, their impact on coding tools, with key community reactions cited
→ 08After GPT-5.5 — Patterns the Community Discovered
Community patterns plus the 2026-04-30 Codex CLI 0.128.0 update — /goal, permission profiles, import, Worktree/Cloud
→ 09Advanced Multi-Agent Codex — Dependency Graphs and Meta-Harness
TaskCreate patterns, TASKS.md-based Codex DAGs, Worktree parallelism, /goal, and self-optimizing meta-harnesses
→Part IV: From LLM Wiki to the AI Scientist
Rise of the LLM Wiki — Karpathy's Obsidian + Claude Code Workflow
The LLM-wiki paradigm Karpathy sparked — what 'put everything in markdown' really means
→ 11My Activity as AI's External Memory — Bidirectional Flow with Obsidian × LLM
Extending post #7 — vault structure, backlinks/tags, how to ask AI, end-to-end setup
→ 12AI Scientist in Practice — AI/Robotics, Bio/Chem/Drug, Autoresearch
Real-world cases by domain, paper search/organize/write workflows; autoresearch as one branch
→