An on-call agent gets one action, then a human
Autonomous systems usually fail through a chain of reasonable actions that add up. The restart didn't help, so scale up, then roll back, and now three things have changed. Let the agent take one action per incident, announced and picked from a fixed list. If that didn't clear the alert, a person decides what happens next.
ai-agentssre
Longer version: the post this came from.