Why Can't Anyone Answer Questions About the Business? — Garrett Galow, WorkOS

TL;DR

• 1

  WorkOS built Studio to kill the back-and-forth on business questions: Instead of non-technical teams waiting on engineers for ad hoc SQL, Studio lets employees ask things like which blog posts drive the most new team sign-ups and get answers or dashboards themselves.

• 2

  The system uses LangGraph plus Claude Opus, but the real trick is runtime context injection: Studio does not stuff every schema and rule into the prompt upfront, it injects tool-specific guidance only when the agent decides to query Snowflake, Linear, or Notion.

• 3

  Reusable widgets make the output reliable: After the LLM helps create a widget, the final artifact is sandboxed JavaScript with UI, APIs, and queries, so refreshing a dashboard reruns code directly instead of calling the model again.

• 4

  Validation matters more than flashy autonomy: Before Studio hardcodes a Snowflake query into a widget, it runs the query and checks that it returns data, because syntactically valid SQL that yields zero rows is still a broken answer.

• 5

  The biggest source of errors is usually business logic, not SQL syntax: Galow says models often miss org-specific filters like excluding deleted entities or requiring active status, so WorkOS encodes those quirks in context blocks once instead of retraining every employee.

• 6

  WorkOS is willing to pay for model quality: Galow says Opus performs enough better than cheaper models that they accept the cost, especially since widget refreshes no longer incur LLM charges.