Multi-agent decisions with a built-in devil's advocate.
Free. Open source. MIT.
Quick Start • How It Works • Protocols • When To Use • Full DNA
We ran the same project through a standard multi-agent session, then through War Room.
| Standard | War Room | |
|---|---|---|
| Features | 10 (over-scoped) | 8 (each justified) |
| Cuts | 0 features questioned | 6 cut (saved 5 dev-days) |
| Risks | Surface-level list | Root cause analysis + switch costs |
| Timeline | "16 days" (optimistic) | "18 days + buffer" (honest) |
| Critical miss | No auto-update | Auto-update moved INTO MVP |
| Alternatives | 0 explored | 3 counter-proposals, best kept as Plan B |
Same model. Same input. Different operating system.
# 1. Initialize a war room
bash scripts/init_war_room.sh my-project
# 2. Write your brief
# Edit war-rooms/my-project/BRIEF.md — what are you building? what constraints?
# 3. Inject the DNA
cp references/dna-template.md war-rooms/my-project/DNA.md
# 4. Tell your agent: "Run a war room on my-project"
# It handles wave orchestration, agent spawning, and CHAOS integration.The agent reads the DNA, picks the right specialists, runs them in waves, unleashes CHAOS after each wave, and consolidates everything into a blueprint.
If you're using OpenClaw:
openclaw skill install war-roomPick 4–13 specialists based on your problem:
| Role | When to use |
|---|---|
| ARCH | System architecture, tech choices |
| PM | Scope, requirements, roadmap |
| DEV | Implementation, code feasibility |
| SEC | Threats, compliance, privacy |
| UX | Interface, interaction design |
| QA | Testing, edge cases |
| MKT | Positioning, launch strategy |
| RESEARCH | Market/tech research, competitive analysis |
| FINANCE | Costs, projections, pricing |
| LEGAL | Contracts, IP, regulatory |
| CHAOS | Always. Non-negotiable. |
Custom roles welcome: AI-ENG, AUDIO, DATA, OPS — whatever the problem needs. See agent-roles.md for the full roster and template.
Agents run in dependency order, not all at once:
Wave 1: Foundation ARCH + SEC + PM → decisions others depend on
Wave 2: Specialists UX + AUDIO + AI-ENG → build on Wave 1
Wave 3: Builders DEV + OPS → implement based on Wave 1+2
Wave 4: Validators QA + MKT + CHAOS → stress-test everything
CHAOS shadows every wave. Not just the end. See wave-protocol.md for the full execution protocol.
The built-in devil's advocate. Attacks every assumption. Rates every decision:
- SURVIVES — withstands scrutiny
- WOUNDED — valid but has weaknesses to address
- KILLED — doesn't hold up, needs rethinking
CHAOS also produces counter-proposals — alternative approaches nobody else considered. In our test run, CHAOS found that 4 of the top 5 failure scenarios came from a single dependency — something the other 12 agents missed.
19 structured decision protocols across 4 pillars. Not suggestions — constraints that every agent must follow.
| Protocol | What it forces |
|---|---|
| Opposite Test | State the opposite decision and argue FOR it |
| Five Whys | Dig to root cause, not surface symptoms |
| Ignorance Declaration | Declare KNOWN / UNKNOWN / ASSUMPTION before analyzing |
| Via Negativa | List 3 things to REMOVE before adding anything |
| Plan B | Every critical decision needs a backup with switch cost |
| Pre-Mortem | "How does this fail in production?" before declaring done |
| CHAOS | Adversarial review of all decisions |
The full DNA adds: Dialectic Obligation, Mirror Test, Ripple Analysis, Tension Map, Causal Chain Verification, Tempo Tagging, Create-Then-Constrain, Barbell Strategy, and Lessons Permanent.
→ Full DNA with all 19 protocols
war-rooms/my-project/
├── BRIEF.md ← Your project description
├── DNA.md ← The operating protocols
├── DECISIONS.md ← Append-only decision log
├── STATUS.md ← Agent completion tracking
├── BLOCKERS.md ← Issues requiring human input
├── TLDR.md ← Executive summary
├── agents/
│ ├── arch/ ← Architecture specs
│ ├── pm/ ← Product requirements
│ ├── chaos/ ← Challenges + counter-proposals
│ └── [role]/ ← Any specialist
├── artifacts/
│ └── BLUEPRINT.md ← Consolidated output
├── comms/ ← Inter-agent messages
└── lessons/ ← Post-mortem learnings
Use it when:
- Decisions cost weeks of work if wrong
- You need multiple perspectives but don't have multiple people
- You need a PRD, architecture, or strategy that survives contact with reality
- You want to stress-test an existing plan before committing
Don't use it when:
- The task is simple and well-defined (just ask your AI directly)
- You need a quick answer, not a deep analysis
- You've already decided and just need execution
Software: "Build a macOS app for AI music generation" → 4 waves, 13 agents, 57 decisions, 32 docs, complete blueprint in 35 minutes.
Business: "Should I pivot from B2C to B2B?" → CHAOS attacks both sides, Five Whys finds the root cause isn't the business model, counter-proposal identifies a third option.
Creative: "Plan the launch strategy for an open-source project" → MKT positioning, RESEARCH competitive landscape, CHAOS finds the distribution strategy is missing.
"Isn't this just prompt engineering?" The protocols ARE structured prompts. That's the point. "Prompt engineering" that produces measurably different results isn't a dismissal — it's a description. The question isn't whether it's prompt engineering. It's whether it works. Run a war room and compare the output to your usual workflow.
"Why would I want AI to disagree with me?" You don't want disagreement. You want accuracy. The CHAOS agent doesn't disagree for sport — it stress-tests decisions so the ones that survive are the ones worth shipping. It's the difference between "great idea!" and "great idea, but here's how it fails."
"19 protocols is too many."
Start with the Essential 7. They cover 80% of the value. The other 12 are there when you want deeper analysis. You don't need to learn all 19 to get started — you need to run init_war_room.sh and write a brief.
"How is this different from CrewAI / AutoGen / MetaGPT?" Those are agent orchestration frameworks — they help you run multiple agents. War Room is a decision methodology that happens to use multiple agents. The difference is the DNA: mandatory protocols that force agents to question assumptions, declare ignorance, and attack their own conclusions. You could implement War Room on top of CrewAI if you wanted.
War Room is open source because the best knowledge is the kind that's passed forward.
Ways to contribute:
- New agent roles — domain-specific specialists
- Protocol refinements — improvements to the DNA
- Example war rooms — complete session outputs for different problem types
- Translations — the DNA in other languages
MIT. Use it, fork it, build on it.
"The unexamined life is not worth living." — Socrates
"Wind extinguishes a candle and energizes fire." — Nassim Taleb
"O melhor conhecimento é aquele que é passado adiante." — Max Kleinz
Created by Max Kleinz / COSMOPHONIX