Workshops Archive

Ops Knowledge That Stops Evaporating: Building an Evergreen Documentation, Search, and Monitoring System on an Agent CLI


Workshop: HPC Systems Professionals Workshop (HPCSYSPROS26)

Authors: Adam Hough (Microsoft, ACM SIGHPC Systems Professionals Virtual Chapter)

Abstract: Operational knowledge in HPC has a half-life measured in shifts. Queries get re-derived, negative results re-litigated, and incident reports cannot be reproduced months later. We describe a small, vendor-neutral system deployed at a large GPU-cluster site, and how to build your own in four weeks. Six pieces, five most sites already have: a terminal-native agent CLI, Model Context Protocol servers exposing telemetry read-only, versioned markdown procedures, a git repository, a local search index, and a few hundred lines of generator code. The load-bearing idea is not the model; it is a write-back rule under which any agent reading the knowledge verifies and extends it in the same task. The same repository distributes it: stable procedures become versioned skills, shipped to the team as a plugin through pull requests. A second-order effect surprised us: when a monitoring page costs a minute, dashboards become disposable. We close with anti-patterns and limits.


Back to HPC Systems Professionals Workshop (HPCSYSPROS26) Archive Listing

Back to Full Workshop Archive Listing