CAVA, the Mediterranean bowl restaurant chain that grew from 133 locations in 2021 to 459+ restaurants in 2026 on the way to a 1,000-location target by 2032, has deployed "Ask Astro" — a multi-agent supervisor built entirely on Databricks that lets its 15,000+ team members ask any data question in plain English and get scoped, governed answers in seconds. The system sits on top of CAVA Core, a single-source-of-truth data platform built on Unity Catalog that already serves 200+ data users, runs 250+ daily jobs across 1,500 tables, and supports 200,000+ AI-assisted queries.
"Can someone pull this data for me?" — every CAVA business team, every week. That recurring request from finance, operations, people & culture, and marketing teams is what the platform was built to solve, replacing the hours-or-days turnaround of analyst tickets, shadow spreadsheets, and dashboard-hunting with sub-minute answers from an internal chat agent that respects each user's Unity Catalog access level so people only ever see data they are allowed to see.
CAVA built the system in two phases. Phase 1 shipped four domain-specific Databricks Genie Rooms — Customer (loyalty, marketing, customer behaviour), Operations (per-restaurant speed of service, catering), People (headcount, turnover, hiring funnel), and Finance (portfolio finance, NRO pipeline, sales rollup) — each giving deep context within its domain but bounded by the Databricks UI. Phase 2 layered Astro on top as a lightweight supervisor that carries enterprise context across tools and surfaces, routing each question to the right Genie space and synthesising a unified answer.
The Astro agent routes through specialised subagents per domain, then composes results and hands back a plain-English answer through an internal chat interface scoped to each user's access level. Each subagent carries tight, scoped prompts and remains individually observable end-to-end. Beyond the Genie spaces, Astro reaches into Atlassian MCP for live Jira and Confluence read/write and into Tavily MCP for web search and page extraction, with the architecture explicitly designed to expand to any available MCP server as new data surfaces are added.
Three layers of observability back the deployment. MLflow Tracing captures per-stage timing, subagent routing, and tool calls as replayable runs. Lakebase stores every user message, assistant reply, thumbs-up and thumbs-down in Postgres, along with end-to-end latency and conversation context. An AI/BI dashboard rolls up cost, latency, activity, and satisfaction by day and by user, giving the team the data needed to iterate on prompts and routing.
The team published five "things to know day one" field notes from the production rollout: identity-aware infrastructure — forwarded user token plus Unity Catalog ACLs plus per-user OAuth — is the real safety seatbelt, not prompt rules; a single missing per-user OAuth consent breaks every question in the agent, so consent CTAs must surface before users hit the wall; supervisor quality caps at the underlying Genie room's 5 QPM API limit; experimentation and evaluation logging from day one is what lets the team improve prompting and routing over time; and parsing must read the last message in a multi-item response because supervisors append "I'll check the data" preambles before the real payload.
Astro is positioned as a build-your-own-bowl for analytics — every team composes its own question across the same governed data products, rather than waiting on a specialist to pull a one-off extract. The team is now extending the same supervisor pattern into additional surfaces, with the Genie rooms, MCP tool catalogue, and observability stack expected to grow as CAVA scales toward its 2030 restaurant footprint.