Reliable data agents need more than database access: they need trusted models, metric definitions, business context, and signals about which sources to use. In a new engineering write-up, the company explains how it rebuilt its internal data stack around that requirement.
The team uses tools including Hex, dbt, semantic models, and observability to make company data easier for agents to query and interpret. LangChain says its data agent now handles about 40 times the request volume that its three-person data team could answer directly.
The important shift is organizational as much as technical. Instead of writing every query or report themselves, the data team now focuses on improving the system: adding guardrails, clarifying definitions, and building feedback loops. The lesson is that agents can expand self-service analytics, but only if the underlying data layer gives them enough context to avoid plausible but unhelpful answers.