Bringing Agents to Data: How Workday Architects Governed Agentic AI with Iceberg and Unity Catalog
Overview
| Experience | In Person |
|---|---|
| Track | Governance & Security |
| Industry | Enterprise Technology |
| Technologies | Delta Sharing, Unity Catalog, Lakebase |
| Skill Level | Intermediate |
Traditional “data-to-compute” models create latency and fragmented governance that hinder Agentic AI. To solve this, Workday and Databricks advocate “bringing agents to data” via the Databricks Lakehouse. By operating on data in situ, autonomous agents maintain real-time consistency without the risks of replication. This architecture enables reliable multi-agent coordination at scale while ensuring unified data integrity and reduced operational costs across the enterprise.
The technical core uses Apache Iceberg for ACID transactions and snapshot isolation, providing a stable foundation for agent reasoning. Databricks Unity Catalog adds a unified governance layer, enforcing fine-grained security, lineage, and auditability. This production-grade approach significantly cuts query latency and storage expenses, allowing organizations to scale from single-agent pilots to hundreds of coordinated agents across multiple departments like Customer Experience, RevOps and Analytics.
Session Speakers
Jason Reid
/Director, Product
Databricks
Phoenix U Majumder
/Senior Director, AI Engg. & Platforms
Workday
Full Summary
Why scaling agentic AI starts with a unified data foundation
Enterprises can prototype agentic AI quickly, but most struggle to scale it with reliability and control. The conversation argues that the difference between momentum and mess is a unified data and governance layer established early, before agents proliferate. The guidance echoes why companies winning in AI build the data layer first: bring agents to well-governed data, not data to scattered agents. Teams looking to put that principle into practice will find that data quality is foundational to any serious AI strategy, since agents are only as reliable as the data they reason over.
FAQ
No. The next iteration of Iceberg will share a unified metadata structure with Delta Lake, making the capabilities discussed applicable regardless of current format. The focus should be on what the storage layer enables for agents, not on the label.
Long-term memory, such as enduring knowledge and agent personas, belongs in the shared Iceberg-based data layer. Short-term session context should live in a vector store to guide reasoning within an interaction. Designing both upfront avoids costly retrofits.
A central governance plane, enforced through Unity Catalog, provides tagging, role-based access control, and lineage. Agents inherit the user's entitlements, so requests outside authorization are blocked, including attempts via prompt injection.
Point-to-point can work for one or a handful of agents, but it breaks at scale. Because agents can be deployed much faster than humans are hired, invest early in centralized data and governance to avoid integration debt and trust failures later.
Three stand out: a unified metadata tree across Delta Lake and Iceberg, native multimodal data types so documents and structured data share lifecycle and governance, and the ability to add columns without rewriting existing Parquet files. These arrive as in-place enhancements rather than migrations.