Red Hat opens the templates behind its internal enterprise data agent

Red Hat has published the architecture behind Dataverse Agent, its internal natural-language interface for enterprise data, and released three open source templates intended to give other teams a starting point for similar systems.

The useful part is not the chat interface. It is the set of controls Red Hat put between a user question and a database query: domain-owned guidance, staged table selection and SQL generation, inherited access controls, and a visible execution trail. That makes the post a practical reference for platform teams deciding what an enterprise data agent needs beyond a model endpoint.

What Red Hat built

According to Red Hat’s engineering account, Dataverse Agent routes a question through four steps: identify the relevant data product, shortlist tables, generate SQL using schema and business guidance, then execute the query. Each step is logged so users can inspect why the agent chose a table and how it interpreted the request.

The system uses Model Context Protocol as the interface between the agent and data services. It runs on OpenShift, authenticates through Red Hat’s existing single sign-on system and inherits the user’s data-platform permissions rather than introducing a separate authorization layer.

Red Hat says the internal deployment exposes 40 official data products to roughly 1,000 unique users, with about 70 daily active users. Those figures are vendor-reported and describe Red Hat’s own environment, not an independent benchmark.

The design choice that matters

The strongest pattern is Red Hat’s “guidance architecture.” Data-product owners maintain structured documentation describing business terminology, valid joins and the intended use of tables. The agent consults that material before generating SQL, reducing reliance on column names alone and keeping responsibility with the teams that understand each domain.

Red Hat also describes an automated intake process that analyzes historical analyst queries to draft guidance for review. That human-review boundary matters: the automation proposes institutional knowledge, but data owners remain responsible for validating it.

What teams can try

Red Hat released three repositories: template-agent for the agent workflow and behavioral configuration, template-mcp-server for Kubernetes and OpenShift-oriented data connections, and template-ui for a React chat interface backed by Fastify. The templates are linked from the original post.

A sensible evaluation starts narrowly: connect one governed data product, test ambiguous business terms, verify permission inheritance and inspect every generated query before execution. The architecture is most credible where teams already have clear data ownership; it does not remove the work of cleaning, documenting and governing the underlying data.