ExaAgent: Trustworthy and Token-Efficient Agentic Analytics of Datacenter Telemetry

The efficiency of large-scale AI and HPC datacenters is limited less by missing telemetry than by who can read it: operational data is partitioned along the facility, system and application pillars of HPC operations, and the cross-pillar questions that actually move utilization and uptime are the ones nobody is equipped to ask. We present \sys, a deployed agentic assistant that answers such questions in natural language over the whole telemetry corpus and returns the plots and statistics an operator needs. \sys couples a planning coordinator with two isolated sub-agents: a retriever whose every data access is a SPARQL query validated against a domain ontology before execution over a Virtual Knowledge Graph (VKG) built per request, and a code sub-agent that runs model-written analysis inside a per-user gVisor sandbox. The design is deliberately restrictive: the LLM never touches a raw table, and we quantify what that restriction buys. On 49.9 TB of production supercomputer telemetry, holding the model fixed, grounding raises end-to-end task correctness from 2/14 to 12/14 and yields a 2.5× token saving on the tasks both configurations solve.