Grafana Labs ships a knowledge graph for every Cloud stack

Grafana Labs announced on October 20, 2026, at ObservabilityCON 2026 in San Francisco that every Grafana Cloud stack now ships with a built-in knowledge graph: a runtime layer that links infrastructure, applications and user-facing signals into one queryable model, so both people and AI agents can trace a symptom back to the dependency that caused it.

What actually shipped

Three pieces went out together. Custom entities, now generally available, let teams group discovered resources — databases, Kubernetes workloads, services — into business-relevant units defined from telemetry rather than hand-maintained inventories. A service catalog with ownership, in private preview, ties each entity to on-call schedules and runbooks. And runtime context for agents, also generally available, exposes the graph through Cypher queries to Grafana Assistant and to third-party agents connected over Grafana's open gcx standard.

Richard Lam, Director of Product at Grafana Labs, framed the problem the release targets, as reported by Yahoo Finance: most teams do not actually have a monitoring gap, they have a technology-debt gap — telemetry pours in from every layer with no durable way to connect it to the customer it affects. A separate reference customer, Neil Wilson of LexisNexis Risk Solutions, put the distinction more bluntly in a separate account of the launch: telemetry says what is happening, context says why it matters.

Why a graph, and not another dashboard

The design choice is the interesting part for anyone who works with structured knowledge rather than raw metrics. Grafana is not adding a new chart type; it is adding an entity-relationship layer that an agent can traverse with a query language built for graphs. That is the same move knowledge-engineering teams have made for years when flat tables stopped being enough to answer why questions across systems that were never designed to share a schema.

It also reframes what "grounding" an AI agent means in production observability: instead of feeding a model more raw logs and hoping retrieval finds the right ones, the agent walks a typed graph of entities and relationships that was already curated by the platform. Teams building this kind of agent-facing data layer into their own products, through GenAI integration work at YuSMP, run into the identical trade-off: a well-typed graph answers cross-system questions a vector search alone will miss, but only if the entity model behind it is kept consistent as the system grows.

The open question

Grafana's entities are discovered from telemetry, which is pragmatic but implicitly schema-light compared with a formally specified ontology. Whether that graph stays coherent as more signal types and third-party agents plug into it — rather than drifting into inconsistent ad hoc categories, the failure mode mid-level ontologies like CCO exist to prevent in data-integration settings — is the detail worth watching as the feature rolls out beyond private preview.