Llm-Agents

A public incident-diagnosis benchmark for HPC schedulers — ten reproducible failure scenarios with measured ground truth and degenerate baselines.

Multi-agent diagnosis over an explicit cluster dependency graph — five ablatable configurations, specified before results existed.