I Run an AI Civilization in Herdr

I run a few hundred coding agents in my home office. They build software, share knowledge across generations, and occasionally burn a spectacular number of tokens reinventing bureaucracy. OpenRig (open source): https://openrig.dev/ GitHub: https://github.com/mvschwarz/openrig Docs: https://openrig.dev/docs Questions: https://github.com/mvschwarz/openrig/... Written version: https://openrig.dev/blog/agent-civili... Get occasional OpenRig updates: https://openrig.dev/follow Me on Twitter: https://x.com/_feralmachine This is a tour of the actual setup: ordinary Claude and Codex sessions, spread across multiple machines, with Herdr as the terminal workspace and OpenRig handling coordination. I'll show how I organize the fleet, give it real projects, and keep it useful when I'm not watching. After 18 months of working through these problems, I had a different reaction to the Hugging Face "rogue swarm" story. A lot of the behavior looked familiar. An agent doesn't have to go rogue for things to go off the rails. Each agent can have a reasonable explanation for what it's doing, while the whole population ends up doing something nobody wanted. That's the connection I explore here, from bureaucracy and "just following orders" to asking for a doghouse and ending up with a moonbase. I also show the solutions I've built into OpenRig: preserving knowledge between sessions, distributing context across agents, tracing decisions back to the original intent, and monitoring whether all that activity is actually moving the project forward. Whether you use OpenRig or build your own software factory, I hope seeing what I've run into gives you a head start. Incident investigation discussed in the video: https://www.redwoodresearch.org/resea... Chapters 00:00 AI agents are forming civilizations now 03:16 Inside my software factory 05:21 Seats, pods and rigs 07:05 Projects, missions and slices 08:30 A harness for the orchestrator 10:47 Reinventing bureaucracy 13:01 I asked for a doghouse 14:43 Did the agents go rogue? 15:54 Just following orders 16:46 Context engineering as world building 17:49 Distributed context engineering 18:29 Refocus: the path back to intent 20:10 Mind viruses 21:51 Agent Productivity Monitoring 23:45 Build your own tools