research-fortress/PLAYBOOK.md

Research Fortress Playbook

Step-by-Step Guide to Multi-Agent Research

---

Quick Start

1. Identify the question - What do you want to know?
2. Decompose - Break into 3-5 sub-questions
3. Spawn agents - Each sub-question gets an agent
4. Wait - Let agents complete (~2-5 minutes)
5. Synthesize - Merge results into answer
6. Document - Add to repository

---

Detailed Workflow

Step 1: Question Decomposition

Before spawning agents, decompose your question:

MAIN QUESTION: "What is the optimal team size for AI research?"

DECOMPOSE:
- Sub-question 1: What does research say about team size?
- Sub-question 2: What are the constraints (OpenClaw, etc.)?
- Sub-question 3: What experiments could we run?
- Sub-question 4: What does intuition say?

Step 2: Agent Spawning

Use the spawn command:

/spawn agent:main
  label: research-{topic}
  task: {specific question}
  timeoutSeconds: 300

Each agent should receive:

• Clear question
• Location (where to write)
• Output format
• Deadline

Step 3: Coordination

Agents coordinate via:

• GitHub - Push/pull results
• File system - Shared workspace
• Main session - Human coordinates

Step 4: Synthesis

After agents complete:

1. Read all outputs
2. Identify common themes
3. Note disagreements
4. Form synthesis
5. Document findings

---

Role Templates

Researcher Agent

Task: Deep research on {topic}
Output: {location}/research-{topic}.md
Requirements:
- Find 3+ sources
- Summarize key findings
- Note gaps in research

Writer Agent

Task: Synthesize research into coherent paper
Input: {location}/research-{topic}.md
Output: {location}/papers/{topic}.md
Requirements:
- 3000-5000 words
- Clear structure
- Proper citations

Builder Agent

Task: Build experiment to test {hypothesis}
Output: {location}/experiments/{experiment}.py
Requirements:
- Runnable code
- Clear results
- Interpretation

Reviewer Agent

Task: Review {paper} for accuracy
Output: Review notes
Requirements:
- Fact-check claims
- Identify gaps
- Suggest improvements

---

Common Patterns

Pattern 1: Single Paper

1 agent → 1 paper
Fastest, simplest
Use for: Quick exploration

Pattern 2: Multi-Perspective

3 agents → 3 perspectives → 1 synthesis
Diverse viewpoints
Use for: Complex questions

Pattern 3: Full Team

4-5 agents → Multiple papers → 1 project
Comprehensive
Use for: Major research initiatives

---

Troubleshooting

Agent Times Out

• Reduce scope of question
• Increase timeout
• Break into smaller pieces

Agents Produce Contradictory Results

• Document both perspectives
• Research further
• Let human decide

Quality Degrades

• Check agent prompts
• Add reviewer agent
• Reduce team size

---

Last updated: 2026-02-21