LangChain's data team shares how they rebuilt their data stack around an agent-first architecture using Hex, dbt, semantic models, and observability. The migration enabled a 40x increase in self-service data request volume, with nearly 100% of provisioned users engaging with the data agent monthly. Key lessons include the importance of rich context layers — dbt column/table definitions, semantic models for metrics, workspace guides for business rules, and endorsements for trust signals — as well as a feedback loop using observability to continuously improve agent responses. The data team's role shifted from answering ad-hoc questions to building and maintaining the context systems that make the agent reliable.
Table of contents
Key resultsWhere we startedHow we evaluated our toolingWhat changed after we migratedHow we think about contextHow we improve the systemWhere we're going nextWhat we’ve learnedClosing368 Impressions