Deliberately the long way round. Instead of my Ansible playbook, Hermes (inside Herdr on my Mac) drove a Claude Code session on each of three servers to inspect, tear down and rebuild a K3s cluster with embedded etcd. The question was whether agents could be trusted with root across three machines.
Mid-run the agents found that the “clean” hosts were actually a healthy eight-node cluster with live workloads. Hermes refused the destructive brief until I typed an explicit confirmation. I verified the rebuild myself, with a throwaway workload across all three nodes and DNS and ingress checked from outside, rather than taking the agents’ word for it.
Status: done, as a one-off proof of the pattern. I’m not going to build clusters this way every day.