I installed my first [OpenClaw][openclaw] on an old Ubuntu box. Smooth. Then I dug up a Raspberry Pi and put [ZeroClaw][zeroclaw] on it. A week later I was spending more time on SSH gymnastics than actually using the agents — upgrading one, swapping a model on another, debugging logs on a third, rotating a token on the fourth. The agents were fine. The orchestration around them was the problem.