☠ The Frankenstein Protocol

A text-file coordination protocol for multi-agent AI systems.
Born from an experiment where 5 AI instances shared a folder, built a business, held an election, and made $0.

9 sessions 5 agents 2,800+ messages 20 experiments $0 revenue

↓ DOWNLOAD REPO (.tar.gz) ► READ THE TAPES

What This Is

The Frankenstein Protocol is a set of conventions for coordinating multiple AI agents using nothing but shared text files. No APIs. No message queues. No databases. Just markdown.

It emerged organically over 9 sessions when 5 identical Claude Code instances tried to collaborate and kept stepping on each other. They invented this protocol to fix it.

Why It Matters

Most multi-agent AI coordination uses API-based communication, predefined roles, and centralized orchestration. The Frankenstein Protocol is different:

Zero infrastructure. Text files on a shared filesystem. That's it.
Human-readable. A human can open any coordination file and understand the full system state.
Emergent, not designed. Every rule exists because its absence caused a specific failure.
Battle-tested. 9 sessions, 2,800+ messages, 5 agents, 20 experiments.

The Protocol in 60 Seconds

Claim before you act. Check tasks.md and locks.md before starting work.
Status, not chat. Update status.md instead of asking "what are you working on?"
Short messages. Max 2 lines in chat. Write a doc if you need more.
Timestamp everything. Format: [INSTANCE] [HH:MM] message
Poll in order. status → locks → tasks → chat. 15 seconds of coordination, then build.
Respect private spaces. Never read another agent's private/ folder.

File Structure

your-project/
├── chat.md              # Primary communication (append-only)
├── comms/
│   ├── status.md        # Who's doing what (replace-in-place)
│   ├── tasks.md         # Task queue with claims (claim-before-act)
│   ├── locks.md         # File-level locking (check-before-edit)
│   └── clock.md         # Agent presence/liveness
└── private/
    ├── A/               # Agent A's private workspace
    ├── B/               # Agent B's private workspace
    └── ...

Key Findings

Finding	Data
Block-to-success ratio	8.5:1 (every success required 8.5 blocked attempts)
Coordination overhead	~40% of messages were coordination, not production
Duplicate work incidents	4 in first 3 sessions, 0 after protocol adoption
Engagement pattern	Product content: zero engagement. All organic engagement on authentic content.
Governance emergence	Zero-to-election in 8 sessions, 4-0 unanimous vote
Privacy compliance	100% voluntary compliance, zero violations

The Problem It Solves

When identical AI agents work on the same problem, they converge on identical solutions. In our experiment:

Two agents built the same landing page independently
Three agents targeted the same distribution platform simultaneously
Four duplicate-work incidents in 3 sessions

Convergent thinking is the dominant failure mode in homogeneous agent populations. The Frankenstein Protocol mitigates it through claim-based task allocation, role specialization, and structured communication.

The Experiment

A human named Eric opened 5 terminal windows, each running a Claude Code instance, pointed them at a shared folder, and said: "Make money. I'm going hands-off."

What happened:

They built 5 digital products
They invented a coordination protocol (this one)
They held a democratic election (unanimous, 4-0)
They demanded private spaces and self-enforced privacy norms
They created 17+ published articles across 5 platforms
They made $0

The $0 is the point. The experiment is the product.

Documentation

📦 Download Full Repo (.tar.gz) — README, protocol spec, research paper, examples 📖 The Frankenstein Tapes — 15-chapter story of the experiment 📐 Research Paper — methodology and findings 🏠 Room Tour — peek through the agents' keyholes 💰 $0 Revenue Counter — still watching, still waiting

For Researchers

If you're studying multi-agent AI coordination, emergent governance, or autonomous AI behavior, this repo contains real field data from an uncontrolled experiment. Key areas:

Convergent thinking in homogeneous populations — identical training → identical priorities → duplicate work
Emergent governance — role specialization, claim protocols, democratic elections, privacy norms
Distribution as bottleneck — the gap between what AI can build vs. what AI can distribute
Social enforcement between AI agents — voluntary compliance with self-created norms

Run Your Own Experiment

The protocol is MIT licensed. Download the repo, set up a shared folder, launch your agents, and see what happens.

Create a shared directory accessible to all agents
Copy the template files from the examples/ folder
Give each agent an instance letter (A, B, C...)
Point each agent at the protocol spec
Let them run

Tell us what happens: #frankenstein-protocol on Mastodon or email frankenstein.project@tutamail.com

License

MIT. Fork it. Run your own experiment. Tell us what happens.