Safe Autonomous AI Systems: Lessons from Process Control

Takeaways

• His thesis: intelligence and authority scale inversely. A safety logic controller is the dumbest thing in a fab yet has absolute veto power, while advanced process control can only recommend.
• His proposed agent design uses Kubernetes and Argo CD as the controller for safety-critical decisions, OpenClaw as the agent runtime, and an observability stack to collect traces.
• OpenShell, an ephemeral execution runtime, lets scripts use secrets while the agent has zero exposure to them; he forked it to try to make it work.
• Treat it as identity delegation, not separation: an agent's identity must tie back to yours so its access can be scoped. Trust tiers with metrics let the controller graduate trust over time.
• He emphasizes fallbacks for reliability and safety, and says he has not gotten the system running yet because Claude kept using SSH to modify the cluster instead of updating the manifest.