MAW — Multi-Agent Workflow for Claude Code & Codex CLI

Prompt rules tell an LLM what to do. MAW adds a verification layer for when it doesn't listen.

MAW is a set of skills that runs a coding task through a chain of independent agents. Each agent checks the previous one's work against the actual code, not against its summary, and assumes the previous agent made mistakes. It complements prompt guidelines like Karpathy's rather than replacing them: it catches what slips through despite the rules.

Approach	How it works	Failure mode
Prompt rules	Tell the agent "be careful"	The agent drifts from the rules under complex context
Self-review	The agent checks its own work	The blind spots that caused the error review the error
MAW	Independent agents verify against the code	Reviewers share no conversation with the author and check the code itself

Quick start

curl -fsSL https://raw.githubusercontent.com/pockerhead/maw/main/install.sh | sh

The installer finds claude and codex and sets MAW up for each one it finds. On Windows, run it from Git Bash. Restart the harness afterwards: skills and subagents are only picked up at session start.

For Claude Code, add the skills to your project's CLAUDE.md:

## Skills
@.claude/skills/maw-execute-task/SKILL.md
@.claude/skills/maw-tasks/SKILL.md
@.claude/skills/maw-context/SKILL.md

Then describe the work and run it:

/maw-tasks "fix 404 on the profile page"   # MAW suggests a mode, you confirm
/maw-execute-task                          # runs the highest-priority pending task

Codex has no custom slash commands, so ask in words: "run maw-execute-task for task 3".

Install flags, upgrades and Windows notes: docs/INSTALL.md. Every command and flag: docs/USAGE.md.

How it works

• Fresh agents. No conversation history crosses a stage. Evidence travels through the task folder.
• Adversarial by default. Every agent assumes the previous one was unreliable. Skeptical, not hostile.
• Premise challenge. Before planning, an isolated agent checks whether the task's premise itself is wrong, and stops for a human if it is.
• Tasks are folders. Status is the folder a task sits in: pending/, in_progress/, done/ or blocked/. Work happens on a feature branch or a git worktree.

Design details: docs/ARCHITECTURE.md.

Modes

Not every task needs the whole pipeline.

Mode	Stages	Use for	Tokens	Cost (Sonnet)
full	Clarifier → Premise Challenge → Planner → Plan Review ×2 → Implementer → Code Review → Fixer → QA	Features, migrations, auth, payments	280–560k	$1–4
small-fix	Implementer → Code Review → Fixer → QA	Bug fixes with a clear scope	120–200k	$0.5–1.5
brainstorm	Clarifier → Premise Challenge → Planner → Plan Review ×2	Exploring approaches before writing code	100–180k	$0.4–1.2
deep-research	Premise Challenge → Planner (web search) → Plan Review ×2	Comparing solutions, auditing approaches	80–150k	$0.3–1

The rule of thumb: if shipping a bug costs more than running the pipeline, use MAW. Opus costs more; Codex stages are billed by your Codex plan. Cost details: docs/USAGE.md.

What you get

Every run leaves an audit trail in the task folder:

maw/tasks/done/TASK-001/
├── TASK_FINAL.md          clarified requirements
├── PREMISE_CHALLENGE.md   premise audit
├── PLAN_FINAL.md          plan after two reviews
├── IMPL_SUMMARY.md        what was implemented
├── IMPL_REVIEW.md         code review findings
├── FIX_SUMMARY.md         what was fixed
├── QA_REPORT.md           test results
└── metrics.md             tokens, tool uses and time per agent

The full list for each mode: docs/USAGE.md.

Any stage on Claude or Codex

Each stage can have its own provider, model and effort. For example, code review on Codex while everything else runs on Claude: a reviewer from another vendor has different blind spots.

{
  "agents": {
    "code-reviewer": { "provider": "codex", "model": "gpt-6.1-sol", "effort": "xhigh" }
  }
}

An invalid model and effort pair stops the run instead of being silently downgraded. The cross-provider path is still experimental. All settings: docs/CONFIGURATION.md.

More skills

• /maw-selfreview: a one-stage review of your own staged diff by a different vendor, fired by a pre-commit gate. Claude Code only. docs/SELFREVIEW.md
• /maw-reason (experimental): adversarial deliberation over a question rather than a task. Five isolated roles end in a synthesis with an explicit disagreement ledger. Claude Code only. docs/REASON.md
• /maw-context: optional project knowledge (invariants, domains, tooling) injected into the agents. docs/PROJECT_CONTEXT.md

What has been measured

Numbers from real use, with their caveats in docs/RESULTS.md:

• A review stage objected in 39.3% of tasks where one returned a verdict (70 of 178).
• The fixer needed a second round in 5.3% of tasks (9 of 169), so review rarely turns into a loop.
• SWE-bench Multilingual: 7 resolved out of 12 instances run, on Sonnet in small-fix mode. There is no same-model single-agent baseline, so this is a data point, not a claim of superiority.
• 15.3% of tasks had at least one spawn die without a verdict (API errors, timeouts, kills).

How it compares

	MAW	Claude Forge	adversarial-dev	Forge AI	adversarial-review (ng)
Approach	Sequential adversarial	GAN-style loops	GAN 3-agent harness	Competing architects + judge	Dual-agent consensus
Scope	Task → QA (full cycle)	Feature + audit + docs	Planning → building → eval	Planning + execution	Code review only
Modes	4 (full, small-fix, brainstorm, deep-research)	brainstorm + audit	—	--fast flag	Cost-gating by diff score
Task management	Built-in (pending/done/blocked)	No	No	No	No
Provider	Claude Code + Codex CLI (per-stage)	Claude Code	Claude SDK + Codex SDK	Claude/Codex/Cursor/API	Claude Code
Install	One curl	cp -r	git clone + pip	pip install	/plugin marketplace

MAW's niche is the full lifecycle with built-in task tracking, mode-based cost control and per-stage providers. If you only need code review, ng/adversarial-review is more focused.

License

MIT