How CoreWeave's Mission Control AI turns Grafana into a live query layer for agents

CoreWeave runs the infrastructure behind some of the world’s largest AI training and inference workloads, and diagnosing issues often means correlating signals across dashboards, metrics, logs, infrastructure resources, and alerts. CoreWeave’s new Mission Control AI brings observability into a conversational workflow, enabling customers and internal teams to ask operational questions in plain English and receive evidence-backed answers grounded in CoreWeave’s live observability data.

In this talk, Senior Engineering Manager Savio Dsouza explains how CoreWeave built Mission Control AI around two custom-built components: an agent harness that plans and executes investigations, and an MCP server that securely exposes CoreWeave’s observability and operational tools to AI clients. Together, they connect the agent to Grafana, Prometheus, Loki, and more.

The MCP layer then encodes CoreWeave-specific observability conventions, infrastructure context, and access controls so the agent can select the right tools, query the right data, and return results with supporting evidence. The harness manages the investigation loop—breaking down questions, gathering evidence across multiple telemetry sources, correlating findings, and producing a grounded response.

Savio also covers how Mission Control preserves the calling user’s permissions and validates tool interactions;  how alert-driven workflows can perform preliminary analysis before an engineer begins an investigation; and the practical challenges of making observability useful to an agent. The result is not simply natural-language access to Grafana, but a foundation for context-aware troubleshooting and operational intelligence.