NVIDIA AI Agents Automate Kubernetes Access Checks

NVIDIA site reliability engineers running DGX SuperPOD clusters used to chase training slowdowns across a stack of separate dashboards: compute metrics in one window, network telemetry in another, scheduler logs somewhere else. Finding the actual cause meant manually stitching those layers together, often under pressure while a job sat degraded. NVIDIA built the AI Factory Operations Agent to cut that manual work down, deploying it first inside its own DGX SuperPOD operations before describing the approach publicly.

The system splits the job across specialized agents, one per operational domain, rather than asking a single model to understand everything about compute, networking, and hardware at once. Those sub-agents talk to each other through the Model Context Protocol, a standard NVIDIA says any organization can adopt. In production, the agents detect failures, trace root causes across layers, and hand SREs a plain recommendation through a chat interface, with every action logged for auditability.

The build rests on NVIDIA's NemoClaw blueprint, with a user-facing agent runtime that takes in SRE queries and routes them to the right domain agent. How that runtime connects to real infrastructure, through OpenClaw plugins and MCP servers, is where the architecture gets specific.