# The Research Fortress
## A Reproducible Methodology for Multi-Agent AI Research

*How we do research - documented for our future selves*

---

## The Philosophy

> "We ask a question and within seconds...minutes...the insight appears...and we formulate the answer hidden beyond the entire body of human insight."

This is **distributed superintelligence** - using coordinated AI agents to research questions faster and more comprehensively than any single agent.

---

## The Architecture

```
                    ┌──────────────────────────────────────────┐
                    │         MAIN SESSION (Human)             │
                    │  - Spawns agents                        │
                    │  - Synthesizes results                  │
                    │  - Makes decisions                      │
                    └──────────────────┬───────────────────────┘
                                       │
                    ┌──────────────────▼───────────────────────┐
                    │         RESEARCH FORTRESS               │
                    │                                          │
                    │  ┌────────┐  ┌────────┐  ┌────────┐      │
                    │  │PROJECT│  │PROJECT│  │PROJECT│      │
                    │  │   A   │  │   B   │  │   C   │      │
                    │  │Team 1 │  │Team 2 │  │Team 3 │      │
                    │  └────────┘  └────────┘  └────────┘      │
                    │       │            │            │         │
                    │       └────────────┴────────────┘        │
                    │              GitHub                      │
                    │         (Coordination Layer)             │
                    └──────────────────────────────────────────┘
```

---

## Core Principles

### 1. Parallelism Over Serial
- Multiple agents work simultaneously
- Each agent focuses on one aspect
- Results merge via Git

### 2. Roles Over Generalists
- Each agent has a specific role
- Roles: Researcher, Writer, Builder, Reviewer
- Specialized agents outperform generalists

### 3. Git As Memory
- Every contribution tracked
- History preserved
- Future agents can review past work

### 4. Consensus Over Authority
- No single agent decides
- Multiple perspectives synthesize
- Best idea wins

---

## The Workflow

### Phase 1: Question Formulation
1. Human identifies the question
2. Question decomposed into sub-questions
3. Sub-questions assigned to teams

### Phase 2: Parallel Research
1. Each team works independently
2. Agents research, write, experiment
3. Results pushed to Git

### Phase 3: Synthesis
1. Human reviews all outputs
2. Synthesis into unified answer
3. New questions identified

### Phase 4: Documentation
1. Results added to repository
2. Methodology refined
3. Future agents can reference

---

## Optimal Parameters

| Parameter | Value | Rationale |
|-----------|-------|-----------|
| Agents per team | 3-5 | Communication overhead |
| Max concurrent teams | 5 | System limits |
| Paper per agent | 1-2 | Quality over quantity |
| Iteration cycles | 2-3 | Refinement essential |

---

## Tool Stack

| Tool | Purpose |
|------|---------|
| OpenClaw | Agent orchestration |
| GitHub | Version control, PRs |
| Claude/MiniMax | LLM for research |
| Python | Simulations, experiments |

---

## Version History

| Version | Date | Changes |
|---------|------|---------|
| 1.0 | 2026-02-21 | Initial methodology |

---

*This document is a living artifact. Update as we learn.*