CAVA's Cava Core data estate, the only data store for the entire company, holds around 3,000 tables. As of about a month before the talk, CAVA had run 200,000 AI-assisted queries year to date, a figure the speakers said was growing exponentially. On top of it CAVA built Ask Astro, a production multi-agent supervisor that serves hundreds of employees with answers in seconds.
“A chatbot generates an answer, whereas a supervisor orchestrates, right?” said a speaker in the session, explaining why Astro is a supervisor rather than a single agent.
Users had traditionally requested data from analysts, often downloading raw files to Excel. Adoption of CAVA's data tools was so high that its small engineering and analytics teams could not field all the questions and were becoming the bottleneck. CAVA had started using Genie about a year earlier and layered new Genie rooms for every business unit, but it had six Genie rooms limited to subsets of data, and the team had to move tables in and out of them every day. That led to the second phase, Astro.
Astro is a lightweight Databricks app built on Dash. When a user logs in, that user's history and context are loaded from Lakebase, where feedback messages, including thumbs up and thumbs down, are also stored. The supervisor, built in Agent Bricks, is a routing LLM that sends each question to domain-specific Genie spaces (five at the time of the talk) or to MCP tools registered in Unity Catalog, then synthesizes the answer. An Atlassian MCP gives access to Confluence, CAVA's document store, and to Jira, so CAVA did not build a RAG pipeline over PDFs, and a Tavily MCP provides internet search. The app passes the user's token so the agent knows who is asking, based on the access controls in Unity Catalog.
For quality control, CAVA runs evaluations every day on the traces Astro produces, using custom scorers for data safety, grounding and routing efficiency among others, and pushes the results to an AI/BI dashboard. It receives an alert when accuracy falls below 85%. MLflow tracing records each step, and an AI/BI dashboard tracks spend, latency and adoption.
The speakers listed lessons from the last three months: enable on-behalf-of-user authorization; provide OAuth consent for every MCP tool, because one missing consent breaks the entire flow; give each Genie room proper instructions, semantics, measures, filters, SQL examples and joins, with one person assigned to each room; enable tracing and keep prompting and routing tuned; and parse the agent's full array of messages so the user sees only the last one. They said a few more MCPs would be added in the next couple of weeks.