Problem
Meta's data warehouse serves analytics, ML and AI. Access used to be handled by teammates and team-level approvals, because data flows stayed mostly inside one team. AI systems and agents now work across many parts of the warehouse, so human decisions struggle to keep up. Managing access gets more complex and takes more time.
Example from the article: a data user woken at midnight to fix a pipeline failure is likely to need data related to that failure, and the system infers this from the user's recent activity.
Solution in brief
Meta split the work into data-user agents, which help people get access, and data-owner agents, which decide requests on behalf of table owners. The warehouse hierarchy is exposed to the agents as text, and the agents infer the business need from user activity. A rule-based risk engine, a daily access budget, query-shape checks and audit logs bound what the agents can allow.
Questions this case answers
- How to automate access approvals over sensitive data with LLM agents while keeping deterministic risk limits and an audit trail
- How to evaluate an LLM decision agent against historical human decisions and catch regressions daily
- How to speed up data-access requests in a large warehouse without loosening security?
- How to let an LLM agent approve or deny access to data and still keep hard limits on what it can allow?
Tools and methods
- Multi-agent system — a data-user agent with three sub-agents behind a triage agent, and a data-owner agent with two sub-agents
- Guardrails — rule-based risk calculation on the agent's output, plus a daily data-access budget
- Human in the loop — people still review permission requests and data-owner decisions
- Golden set — daily regression check on real historical requests
Google or a one-time email code. No password.