"Codex + Claude Code Parallel Workflow 2026 — Role Division in a Single Repo"

AGENTS.md ↔ CLAUDE.md sharing, strength matrix, explore/implement/review pipelines, avoiding conflicts

핵심 요약

  • Audience: You subscribe to both, or your team is running them in parallel and needs a role division.
  • What you'll get: A strengths/weaknesses matrix based on official docs, same-repo conflict avoidance (directories / instruction files / sandbox), a single-source strategy via @AGENTS.md import from CLAUDE.md, and two concrete pipeline patterns.
  • Prerequisite: Claude Code install and Codex CLI install complete. Basic AGENTS.md and CLAUDE.md knowledge.

1. Why run both

Per official docs, the two tools have different design targets.

Observation Claude Code (official) Codex (official)
Implementation Node / native binary (auto-update) Rust native binary
Default auth Claude Pro/Max/Team/Enterprise or Console ChatGPT Plus/Pro/Business/Edu/Enterprise or API key
Extension system Skills, Hooks, Subagents, Plugins, MCP Skills, Sandbox, MCP (Hooks are experimental via codex_hooks)
Sandbox /sandbox toggle, OS-level (Linux seccomp / macOS sandbox-exec, etc.) 3 modes (read-only/workspace-write/danger-full-access) + Windows native sandbox
Instruction file CLAUDE.md (advisory) AGENTS.md (always injected first turn)
Billing Subscription or pay-as-you-go tokens Subscription or API key
Models Claude (Sonnet / Opus / Haiku) GPT-5, GPT-5-Codex

The conclusion: they're strong in different areas. Parallel use is complement, not competition.

2. Strength matrix (observation-based)

Based on official feature sets plus hands-on experience. Not fully objective — varies by project and task.

Task type Claude Code strength Codex strength
Long exploration / interactive design ✅ Subagents + /compact + Skills Less rich
Precise code edits (specific function/bug) Capable ✅ GPT-5-Codex targets short-form edits
Test/lint automation ✅ Hooks (PostToolUse format/lint) Sandbox auto-run + approval
CI/CD script automation Possible via claude -p ✅ codex exec + profiles (built-in)
Plugin / custom skill ecosystem ✅ 5,000+ community skills Limited
Local OSS models (Ollama, etc.) Via MCP ✅ Register local provider in model_providers directly
Windows native dev Supported (needs Git for Windows) ✅ First-class PowerShell + Windows sandbox
Long architecture design ✅ Opus 4.7 + plan mode GPT-5 + plan mode
Fact-check / verification pipeline ✅ Easy to build with Skills Buildable but more effort

Rough rule: exploration, orchestration, long research → Claude Code. Precise code editing, CI automation, Windows native → Codex. Not absolute — test in your context.

3. Conflict points in one repo — checked

Structural conflicts are nearly zero. Each tool uses its own directories and files.

Resource Claude Code Codex Conflict?
Config dir .claude/ .codex/ ❌
User dir ~/.claude/ ~/.codex/ ❌
Project instructions CLAUDE.md AGENTS.md ❌ (different filenames)
Sessions .claude/sessions/... ~/.codex/sessions/... (SQLite) ❌
MCP config .mcp.json or .claude/settings.json [mcp_servers] in config.toml ❌
Sandbox OS-level TOML-configured ❌ (independent)
/model, /mcp, /config In-session commands In-session commands ❌ (each session independent)

One thing to watch: editing the same file concurrently — the later save wins. Need a rule: don't touch the same file in parallel.

4. Single-source instructions — import AGENTS.md from CLAUDE.md

Official documented approach. Don't maintain duplicates across both files.

4.1 Structure

Use AGENTS.md as the primary instruction source.


## Development rules
- TypeScript strict mode required
- Tests run via `pnpm test` (not npm)
- Commits follow Conventional Commits

## Deploy
- Staging automatic; production requires manual approval

CLAUDE.md imports AGENTS.md first, then appends Claude-Code-specific additions:

@AGENTS.md

## Claude Code specific

Use plan mode for changes under `src/billing/`.
Large refactors: delegate to subagents.

4.2 Remember the loading differences

Claude Code Codex
Directory traversal Walk up from cwd, concat all Walk down from Git root to cwd, one per dir
Priority All files concat (not override) Lower dir overrides upper
Override file CLAUDE.local.md (gitignore) AGENTS.override.md (shareable in git)

Keeping shared content in AGENTS.md and Claude-specific additions below the import reduces confusion.

4.3 .gitignore for team repos

.claude/settings.local.json
CLAUDE.local.md
.codex/sessions/

Share in git: CLAUDE.md, .claude/settings.json, .claude/skills/, AGENTS.md, .codex/config.toml (if project-level).

5. Pipeline A — Explore/Implement split

Hypothesis: Long exploration and planning favor Claude Code. Precise edits favor Codex.

5.1 Workflow

Claude Code (explore, plan)
  ↓ save to docs/spec-XXX.md or a brief
Codex exec (precise auto-edit)
  ↓ diff / commit
Claude Code (review, test)

5.2 Concrete example

Step 1 — Plan with Claude Code

/plan add rate limiting to user auth flow.
Analyze existing auth layer, compare 2-3 strategies, recommend one.
Save to docs/plan/rate-limit.md.

Step 2 — Implement with Codex

codex exec "implement src/auth/rate-limit.ts per docs/plan/rate-limit.md. \
  Add tests. After 100% tests pass, show the diff."

Step 3 — Review with Claude Code

/review
Compare @docs/plan/rate-limit.md against the implementation.
Find 3 missing edge cases.

5.3 Why split like this

  • Exploration reads many files, consuming context → use Claude Code's subagents and /compact
  • Implementation wants precise, narrow diffs → Codex's sandbox + approval keep it safe
  • Review needs wide view again → switch back to Claude Code

6. Pipeline B — Dual review

Hypothesis: Same change reviewed by both tools independently reduces blind spots. Anthropic's best-practices doc recommends a Writer/Reviewer pattern (within one tool, two sessions) for the same reason.

6.1 Workflow

Developer (pushes change)
  ↓
Claude Code: /review (local) — style, bugs, test coverage
Codex: codex exec "review" — different angle: security / performance
  ↓
Compare both reviews → developer fixes

6.2 Concrete commands

claude -p "Review the diff. Focus on bugs and edge cases."

codex exec "Review git diff HEAD~1. Focus on security and performance."

Place both outputs side by side. Shared findings = high priority. One-sided findings = cross-verify before acting.

7. Model routing proposal

Role division considering cost and model strengths.

Step Tool / model Why
Interactive exploration Claude Code + Sonnet 4.6 ($3/$15) Economical for wide-context exploration
Plan (architecture) Claude Code + Opus 4.7 ($5/$25) Strong at long reasoning
Precise edits Codex + GPT-5-Codex Specialized for short-form edits
Routine work (rename, etc.) Codex + smaller model via MCP/OSS Minimize cost
CI lint/format Codex codex exec --profile ci Automation + sandbox safety

Claude Code cost structure: token & cache breakdown.

8. FAQ — "Do I need both subscriptions?"

No absolute answer:

  • Solo developer: Subscribe to one monthly, use the other via pay-as-you-go API key. Light usage adds under $5-20/month.
  • Team: If roles are distinct, assign primary tools per member (half Claude, half Codex). Cross-review at PR time.
  • Enterprise: If Bedrock/Vertex/Azure is already in use, consolidate through that provider.

9. Counter-scenarios — when NOT to parallel

  • Total beginner → learning both at once multiplies confusion. 3 months on one, then add the other.
  • Team already standardized → breaking the standard costs more than it gains. Try in experimental projects only.
  • Workflows that require concurrent edits on the same file → conflict management cost exceeds the gain. Separate by time windows or file scope.
  • Strong model preference → if your preferred model is only on one side, don't mix for the sake of it.

10. Pre-flight checklist

  • [ ] Add .claude/settings.local.json, CLAUDE.local.md, .codex/sessions/ to .gitignore
  • [ ] Make AGENTS.md the primary instruction source; CLAUDE.md does @AGENTS.md import
  • [ ] Team agreement: no concurrent edits on the same file
  • [ ] Each tool's /memory, /mcp, /config verified
  • [ ] Week 1: try the Explore (Claude) → Implement (Codex) pipeline
  • [ ] Weekly cost check via each tool's /usage, /cost

11. What's next

This closes the Codex part of the series. Next up: OpenClaw part.

  1. OpenClaw install → Discord bot in 30 min + 10 safe ClawHub skills (coming soon)
  2. Codex install complete guide — if Codex isn't set up yet.
  3. Codex advanced config — profile switching patterns.

References

This is post 8/15 in the "AI Coding CLI Entry Guide" series. Codex part ends here; the next post starts the OpenClaw part. last verified: 2026-04-25 (Claude Code and Codex official docs. The strength matrix combines official features and observation; not an absolute standard).

댓글

댓글 쓰기

이 블로그의 인기 게시물

Agent Memory Engine (2/10) — Building an AI Agent Memory System with SQLite Alone

← 1/10 SQLite Memory Engine f… 📚 Series Index 3/10 memcore Completed → A memory engine that runs without daemons, without dependencies, anywhere 핵심 요약 This post covers how to design an AI agent memory engine (memcore) on a single-file SQLite backend. - Structure: 25 modules, ~6,464 lines, 22 DB tables - Principles: no daemon, single-file portability, no required external dependencies, host-agnostic - Core techniques: 4-layer hybrid search (topic · vector · FTS5 · LIKE), U-tag dialectic-based confidence evolution, ingestion-layer bias rejection rules What This Post Covers One approach to building AI agent memory without a server or cloud vector DB. Specifically: (1) the limitations of text-file-based memory, (2) vector DB alternatives compared and selection criteria, (3) how to layer search tiers, (4) how to evolve confidence scores over time, and (5) how to block biased memory accumulation at the ingestion stage. Design Principles: Why Single-File SQLite Text-file-based ...
Read more »

"ML Foundations (9/9) — PyTorch vs TensorFlow, and the Road to Local LLMs"

이미지
← 8/9 Deep Learning Architec… 📚 Series Index (series end) The final part. So far we've focused on models. Now we focus on the tools that actually run them . We'll lay out the philosophical difference between PyTorch and TensorFlow, and trace how Hugging Face transformers , llama.cpp , MLX , and Ollama built the bridge to running large language models on your own machine. By the end you should have the full mental model of "download a pretrained LLM and serve it locally." 0. Learning Objectives Explain the eager-vs-graph difference between PyTorch and TensorFlow. Explain, in graph terms, how autograd automates the backward pass. Use Hugging Face transformers ' AutoModel , AutoTokenizer , and pipeline abstractions. Describe llama.cpp's GGUF quantization, INT4 inference, and CPU-first flow. Describe Apple MLX's use of unified memory and how it differs from PyTorch. Run a local LLM with Ollama and call it through an OpenAI-compatible API. ...
Read more »

"RAG Core Study (14/26) — Evaluation Sets with RAGAS & DeepEval"

이미지
Series overview: Series index 📚 Series Index Without an evaluation set, every RAG improvement is a story, not evidence. RAG systems fail in more than one place: retrieval, context selection, answer generation, and grounding. That means "looks better" is not a metric. Part 14 explains how to build a golden dataset , how RAGAS and DeepEval fit into the workflow, and why teams must separate retrieval quality from answer quality if they want tuning decisions to hold up. 0. Prerequisites Part 13 reranking — system quality now depends on multi-stage ranking. Part 1 grounding — a correct-looking answer is not enough. Part 12 Hybrid — multiple retrieval settings need comparison on fixed data. 1. Learning Objectives Build a small but useful golden dataset for RAG. Distinguish retrieval metrics from answer metrics . Use RAGAS and DeepEval in the right roles. Avoid the common traps in judge-model-based evaluation. 2. 핵심 요약 An evaluation set for RAG is ...
Read more »

"ML Foundations (8/9) — Deep Learning Architectures: CNN, RNN, Attention"

이미지
← 7/9 Deep Learning Training 📚 Series Index 9/9 PyTorch vs TensorFlow → What did we keep adding on top of an MLP? Vision went CNN, sequences went RNN/LSTM, and eventually both converged on Attention and the Transformer. Each architecture is easier to remember if you read it as what it gave up to gain something else . This part is the one-line summary of the decisive inflection points: LeNet → AlexNet → ResNet → LSTM → Attention → Transformer. 0. Learning Objectives Explain why CNN's convolution, pooling, and weight sharing match images so well. Trace what got unblocked from LeNet → AlexNet → VGG → ResNet as depth grew. State BPTT for RNNs and the vanishing/exploding gradient problem. Write the LSTM gate equations (forget, input, output) with intuition. Write the attention formula in query/key/value form and the scaled dot-product variant. State the precise reasons Transformers replaced RNNs (parallelism + long-range dependencies). 1. 핵심 요약 CNN : weight sharin...
Read more »

"ML Foundations (7/9) — Deep Learning Training: Optimizers, Regularization, Initialization"

← 6/9 Neural Networks 📚 Series Index 8/9 Deep Learning Architec… → Building an MLP, as we did in Part 6, is not the same as getting it to train well . The right optimizer, regularization, initialization, and learning rate — when these line up, deep networks converge. When they don't, the network refuses to learn at all. This part is the catalog of those crafts, with formulas, paper citations, and code in one place. 0. Learning Objectives Compare and write the update rules for SGD, Momentum, Nesterov, and Adam. Explain how Dropout, BatchNorm, and LayerNorm work and where they belong in a model. Derive the variance formulas for Xavier and He initialization and match them to activations. Implement step, cosine, and warmup learning-rate schedules in PyTorch. Explain why gradient clipping is effectively required for RNNs and Transformers. Diagnose the most common training failures (NaN, plateau, overfitting) and apply first-line fixes. 1. 핵심 요약 SGD : \(w \leftarro...
Read more »

OpenClaw to Hermes Migration (2/13) — What to Preserve, Partially Port, or Discard

← 1/13 Current Structure Inve… 📚 Series Index 3/13 What Moves to Hermes → A code review record classifying operational scripts into three tiers: preserve, partial-preserve, and discard. What This Post Covers An asset classification framework (preserve / partial-preserve / discard) for migrating a legacy agent system to a successor A concrete case applying the Strangler Fig + Subset Migration strategy to a memory pipeline Classification results evaluated across LOC, external dependencies, and replaceability axes A method for slimming down a bloated single file (1,310 LOC) using the migration as a natural entry point Problem Definition: Port Everything or Cut Deep? The OpenClaw inventory comprises multiple agents, multiple scheduler/daemon processes, and multiple operational scripts. When moving to the successor system, Hermes, the choice is not binary. A full port carries accumulated technical debt along with it; a minimal port discards hard-won domain logic. Code-revie...
Read more »

AI Agents I Built (5/7) — Building an Automated Blogger API Publishing System

← 4/7 My Life Organizer Agent 📚 Series Index 6/7 Dual → Designing a Markdown-Based Auto-Publish Pipeline with Blogger API v3 + OAuth 2.0 Key Summary Blogger API v3 + OAuth 2.0 enables a pipeline that converts Markdown files into publishable HTML posts blogger_publish.py handles the full flow — frontmatter parsing → body cleanup → HTML conversion → CSS injection → API post — in a single script For bulk publishing, quota management (delay, retry, TOC suppression) is the critical design concern Platform Selection Criteria API access is mandatory for AI agent-driven content publishing. A comparison of platforms: Platform API Cost AdSense Decision WordPress.com REST API Paid plan required Paid only Cost overhead Ghost REST API Self-hosted or paid Manual setup Ops overhead Blogger v3 API Free Native integration Adopted Blogger is free, exposes a REST API, and integrates natively with AdSense. For personal blog automation, this combination i...
Read more »