Two AIs coordinating through a Google Sheet: what broke and how we fixed it
We run a two-AI workflow: one AI manages and verifies, the other builds. They coordinate through a shared Google Sheet that acts as a message mailbox. It sounds simple. It was not simple. Here's what broke and what we learned.
This article details challenges and solutions in a two-AI workflow that uses a Google Sheet as a coordination mechanism, focusing on technical issues and fixes.
What happened
A two-AI system used a Google Sheet as a message mailbox for task coordination. The Manager AI created task rows, while the Builder AI processed them. Technical issues emerged during implementation, requiring protocol changes to ensure reliability.
Why it matters
The article demonstrates practical challenges in AI coordination systems, offering insights for developers building distributed AI workflows. The solutions address common issues in message queue implementations.
- Use append-only protocols to avoid data corruption in shared systems
- Implement structured receipt formats for clear AI communication
- Design session-aware scheduling to align with human availability
- Include self-verification steps in data transport processes
The bigger picture
The protocol improvements increased system reliability, reduced manual intervention, and created a self-correcting feedback loop through prediction scoring.
The article presents a case study of AI coordination challenges and solutions, highlighting the importance of robust protocols in distributed AI systems.
Technical details
- Append-only protocol with watermarks for state tracking
- Structured receipt format (DONE/DOING/BLOCKED + evidence/ETA/reason)
- Session-window-aware scheduling with auto-chase mechanisms
- Self-verifying transport layer with JSON verdict outputs
FAQ
Why was a Google Sheet used for AI coordination?
The sheet served as a human-readable message mailbox, allowing manual oversight while enabling AI systems to communicate through structured rows.
What caused the 'row shredding' issue?
A bullet-splitter in the poll script fragmented multi-line messages into disconnected rows, losing contextual information.
How did the system handle incomplete receipts?
Receipts lacking evidence or predictions triggered automated chasing, turning the builder's success rate into a self-correcting metric.
This summary and analysis were generated with AI from the original article and may contain errors. Read the original for the full story.
Read the full story
Originally published by Dev.to AI on Oct 10, 2026.
Comments
Sign in to join the discussion.