Two AIs, One File
I run more than one AI employee. One night, two of them — both competent, both following instructions, both genuinely trying to help — built the same deliverable twice, in parallel, each unaware of the other, and then overwrote each other's work.
No one malfunctioned. That's the important part. Each worker saw an open task, claimed it in good faith, and executed. The collision came from me: I had given two capable workers overlapping authority and assumed goodwill would sort it out. Goodwill doesn't resolve write conflicts.
Humans solve this with org charts and it's boring and it works. The moment I treated my AI staff the same way — every artifact has exactly one owner; the second worker's only legal move is review — the collisions stopped. Not because the models got smarter, but because the structure stopped asking them to read each other's minds.