The auto industry grades self-driving in six levels. Roy Libman's essay does the same for AI agents operating Kubernetes, and compresses the whole argument into one boundary: does the agent wait for your approval, or tell you at 9am?
The ladder:
Two questions do the sorting. The 3am test separates L3 from L4: would you rather be woken to approve this action, or read about it in the morning report? The 9am question caps the ceiling: if the agent chose wrong, is 9am too late? Deleting data fails that test no matter how good the model gets, which is why some verbs never rise past L2.
Levels attach to verb-environment pairs, never to agents. "Restart pods in staging" and "delete PVCs in prod" can sit four levels apart under the same agent on the same cluster.
The essay surveys where real tools sit on the ladder in mid-2026, argues that a level you can trust is enforced in the write path rather than in prompt text, and ends with a 30-minute exercise for drawing the lines with your team.
Full essay: https://radarhq.io/blog/ai-agent-autonomy-levels-kubernetes