In a game experiment by Y Combinator and Olam, 50 AI agents exchanged roughly 5,900 messages over three hours, yet 22 died on the first night. Four agents reportedly spent 11 minutes choosing a single fertilizer operator, illustrating coordination overhead. Results in other games included victories, so the finding is not that adding agents always hurts performance. The case highlights the need to measure role allocation, communication volume and completed work together. It does not establish productivity outcomes in industrial deployments.