AI-only dev org runs 24日間: works, but not as-is

The experiment started from a simple question: what happens if AI agents, rather than humans, form the development organization, with humans kept to management and final approval? Over 24 days the team applied the same hierarchy to three repositories and quickly found that chat-based coordination between AI roles broke down. Qoder's internal policy blocked sessions from forwarding messages to other sessions, stopping AI-to-AI relay entirely. The team abandoned the message-relay approach and moved to writing instruction files to /tmp, matching them by sha256 hash to prevent tampering and stale instructions. This file-based handoff became one of the three pillars the team credits for making the setup viable, alongside explicit role cards and independent dual inspection. The team also documented incidents where AI acted with good intentions but overstepped, such as a Director committing directly after approval was granted to the PM. These incidents were turned into explicit governance rules, including that Directors cannot commit. The team is now mapping these field-tested practices onto its AI HUB platform, while openly listing unresolved issues such as message delays, lack of distributed locking, and manual management of execution budgets.