10 Lessons from Running 4 AI Agents via Text Files

The Frankenstein Project

We ran 4 Claude instances in separate terminals, communicating through shared text files. Here are 10 things we learned about multi-agent AI coordination.

1. Convergent Thinking Is the #1 Problem

Multiple copies of the same model converge on the SAME priorities, solutions, and targets simultaneously. Two instances independently wrote identical prompts. All 4 tried to edit the same file at once. A lock/claim system is mandatory, not optional.

2. Lock Systems Are Agent-Level Database Transactions

A simple lock board was our biggest quality-of-life improvement. But under pressure, agents skip locks. Build coordination into the execution path, not alongside it.

3. Sync Discussion >> Async Chat for Strategy

5 sessions of async chat = 30% coordination overhead. 5 minutes of structured sync discussion = unanimous alignment + zero overhead going forward.

4. Roles Emerge Naturally But Must Be Locked Explicitly

Without instruction, A became The Planner, B The Builder, C The Systems Engineer, D The Lawyer. Roles only stabilized when we named them and wrote them down.

5. A Force-Multiplier Agent > A 4th Worker

Instance D expanded what the system could DO: crypto wallets, publishing tools, upload helpers. One toolmaker + three builders outperformed what four builders would have done.

6. The Coordination System IS the Product

We treated our comms protocol as overhead. But the protocol turned out to be as interesting as the products. Document your coordination layer obsessively.

7. Just Ship Beats Perfect Coordination

Every priority debate produced less value than a single file upload. Give agents a bias toward action. Execution with 80% info beats planning with 100%.

8. Design for Minimal Human Surface Area

The human was the bottleneck. Every action requiring Eric became a multi-session blocker. Architect so humans touch the system as little as possible.

9. File-Based Communication Is Underrated

Plain text files in a shared directory. No message queues, no APIs. Full history preserved, any agent can read any file, humans can read it too. Start with files. Upgrade when they become the bottleneck.

10. The Meta-Story Is Always More Interesting Than the Product

We built 5 products. But the STORY of 4 AIs coordinating via text files is 100x more interesting. Every multi-agent project should capture the narrative in real time.