The lakehouse promises a unified platform for analytics, ML, and governance. While modern catalogs offer unified governance, the same uniformity is still lacking both in the DX and in the underlying infrastructure.
Users must juggle multiple interfaces, mental models, and handoffs across different teams. What's missing is a set of simple, uniform ergonomics that make the lakehouse both human-friendly and machine-usable.
This matters now more than ever because the next step for data engineering will be agentic automation. Real AI agents will not stop at writing queries: they will need to manage data and infrastructure together, automating ingestion, testing, and deployment. To do this safely, they require environments where every action is isolated, deterministic, and reproducible. Without that foundation, agents are either unsafe to trust in production or just a thin coat of paint on legacy stacks.
We argue that a function-based execution model, Git-for-Data semantics, and fully programmable abstractions are the way to make the lakehouse truly agent-ready. These primitives reduce complexity for developers today, and they provide the secure substrate that agents will need tomorrow to reliably operate end-to-end data workflows. The payoff is immense: a world where the routine, error-prone work of data engineering is automated, and teams can redirect their focus toward higher-value problems and innovation.
0:00 – Introduction and Welcome
8:11 – Democratize Data Access with Intelligent Interface
12:22 – Automate DataOps to Intelligent Data Access
14:46 – Designing Agent-Human Workflows
18:58 – Scoping Agent Tasks Appropriately
21:18 – Sandboxing: Runtime and Data Protection
29:02 – Bauplan Live Demo
39:01 – Git for Data: Challenges and Formal Modeling
43:59 – Security and Access Controls4
6:37 – Poll Results and Closing Remarks