Data Agent

Custom Build. Scoped per engagement.

Anyone on your team asks a business question in plain language, in Slack or on the web. The agent finds the right tables, writes the SQL, checks it, runs it under that person’s permissions and answers with the query and its sources in plain view.

It’s the pattern behind OpenAI’s in-house data agent, built on your warehouse. The model is the easy part. The value is in your schema, your metric definitions and your access rules, and that’s the work we do with your data team.

BYOC (bring your own cloud): it deploys to your AWS, GCP, Azure or on-prem environment and queries your warehouse where it lives.

What we build

• Context layer: table descriptions, join keys, metric definitions and a business glossary, organized by domain
• Governed access: an MCP server over your warehouse or semantic layer, enforcing row- and column-level access as the person asking
• A SQL loop that checks itself: generate, then validate (dry run, cost limit), then fix, before anything runs for real
• Answers you can audit: every answer shows its SQL and the tables it came from
• Accuracy tracking: a golden-question set so you know when the agent gets better or worse

Tools / stack

• Claude Agent SDK: the agent loop
• MCP: governed access to Snowflake, BigQuery, Databricks or Postgres
• dbt / semantic layer: reused if you already have one
• Slack app + web UI: where your team asks

When it fits

Your data team spends its week answering the same “can you pull…” requests, while everyone else waits. You want self-serve answers without handing out raw warehouse access.