research-fortress/METHODOLOGY.md

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.