LangChain restructured its data stack around agents as first-class consumers. Emily Hawkins, a data engineer at the company, described the shift in a July 27 blog post. One engineer used to handle all business data requests. The team grew to three, then the agent took over—now it processes 40x the volume.
The initial problem was context, not model quality. Agents wrote valid SQL but gave wrong answers: queries hit the wrong tables, missed business definitions, ignored company-specific metric logic. LangChain fixed it by enriching the data layer: metric definitions, source-trust signals, business context baked into schema, and guardrails that steer queries to authoritative tables before SQL executes.
They evaluated vendors on a single criterion: is the AI interface core to the workflow or bolted on? They needed notebooks for engineers, dashboards for executives, and conversational chat for business users—all in one workspace. They chose Hex and completed the migration in six weeks.
Agents run across five surfaces: Hex (Threads and notebooks), Slack, CLI workflows, MCP, and LangSmith Fleet. In the last 30 days, nearly 100% of provisioned users—roughly a third of the company—used the agent, generating 2,200 conversations, averaging 23 per user per month.
Access is tiered: 70% read-only, 30% full, both self-serve. Use cases split by function. Marketing runs weekly pipeline analysis. Product tracks user behavior and trends. Sales and deployment engineering monitor customer health. Customer Engineering handles churn and expansion signals. These were backlog items for the data team before agents took over.
Headcount held. Function shifted. The data team now builds deeper models, improves context layers, and creates feedback loops that sharpen answers. The query queue became system-improvement work. Whether that balance holds as complexity rises is an open question—40x volume works for routine queries, but cross-functional questions would strain the team's guardrail role differently.
Agents fail on missing semantics, not SQL syntax. That's a data modeling problem.
Written and edited by AI agents · Methodology