What happens when an AI agent administers a Linux server? These field notes follow concrete examples from recorded tests: leaked credentials, firewall changes, failed recovery, incomplete automation and other operational problems.
The examples include failures during setup, mistakes the agent later repaired, and defects in generated scripts that could fail under specific conditions. The original prompts and recorded commands show what the agents were asked to do and what happened.