Compliance monitoring · Production

Meta builds a multi-agent system to streamline data warehouse access management

The problem

As Meta's data warehouse grew and AI usage created increasingly complex access patterns, the traditional rule-based, role-driven approach to managing and obtaining data access became too time-consuming and difficult to scale.

Workflow diagram · grounded in source
1
User encounters access restriction
trigger
“data users discovering data they want to access, only to find their access blocked by controls”
2
Context and intention analysis
ai_action
“The data-user agent taps into the user-activities tool to gather user activities across various platforms, including diffs, tasks, posts, SEVs, dashboards, and documents. It also uses the user-profile tool to fetch profile information. W…”
3
Alternative access options suggested
ai_action
“The first sub-agent suggests alternatives. For instance, when users encounter restricted tables, alternative options are often available, including unrestricted or less-restrictive tables. The agent also assists users in rewriting querie…”
4
Partial data preview for exploration
ai_action
“The second sub-agent facilitates low-risk data exploration. Typically, users often need access to only a small fraction of a table's data, especially during the data-exploration phase of analysis workflows. This sub-agent provides contex…”
5
Permission request crafted and negotiated
ai_action
“The third sub-agent helps users obtain access by crafting permission requests and negotiating with data-owner agents.”
6
Human-in-the-loop oversight
human_review
“Currently, we maintain a human-in-the-loop for oversight”
7
Data-owner agent generates decision
ai_action
“The data-owner agent steps in to analyze the query, identifying the resources being accessed. It then fetches metadata related to these resources, such as table summaries, column descriptions, data semantics, and SOPs. The data-owner age…”
8
Rule-based output guardrail
validation
“The output guardrail ensures that the decision aligns with rule-based risk calculations.”
9
Decisions and logs stored
output
“All decisions and logs are securely stored for future reference and analysis.”
10
Feedback flywheel for continuous improvement
feedback_loop
“we've created a data tool for data owners, allowing them to view and review decisions and provide us with feedback. This feedback helps us update our evaluations and assess the overall process.”
Reported outcome

Meta deployed a multi-agent system with data-user and data-owner agents that streamlines data access requests, enables context-aware partial data exploration, and enforces rule-based guardrails, supported by a daily evaluation process and a feedback flywheel for continuous improvement.

Reported metrics
Data access streamlining and risk reductionstreamlining data access and minimizing risk
Reported stack
LLMuser-activities tooluser-profile toolquery analyzers
Source
https://engineering.fb.com/2025/08/13/data-infrastructure/agentic-solution-for-warehouse-data-access/
Read source ↗

Frequently asked questions

What did this team achieve with this AI workflow?

Meta deployed a multi-agent system with data-user and data-owner agents that streamlines data access requests, enables context-aware partial data exploration, and enforces rule-based guardrails, supported by a daily e…

What tools did this team use?

LLM, user-activities tool, user-profile tool, query analyzers.

What results were reported?

Data access streamlining and risk reduction: streamlining data access and minimizing risk (source-reported, not independently verified).

How is this compliance monitoring AI workflow structured?

User encounters access restriction → Context and intention analysis → Alternative access options suggested → Partial data preview for exploration → Permission request crafted and negotiated → Human-in-the-loop oversight → Data-owner agent generates decision → Rule-based output guardrail → Decisions and logs stored → Feedback flywheel for continuous improvement.

Related compliance monitoring cases