Scenario: an AI is given root permissions on a small cluster (10 - 100 machines). It is given natural language instructions with minimal technical content by a human manager (e.g. it can be told "setup slurm on the cluster" but not "download these links, run these bash scripts on each machine, ...").

For any issues involving physical issues with the cluster, the AI must be able to walk a coworker through a fix process (running a robot that fixes the issue also counts, but is not required).

The AI can be specialized to work with a particular stack (e.g. if it only knows how to use one job scheduler I'll accept that)