What Is an Agentic Lakehouse?
Why a lakehouse designed for human analysts falls short for AI agents, and the four layers that close the gap.
- Agents inherit none of your context
- Four layers over open storage
- An answer you can defend
Video Explainers
8 short explainers on the pieces an AI agent needs before it can be trusted with your data: governed access, business context, a protocol to talk through, versioned tables, and a safe way to write.
Each one runs about 35 seconds at 1080p. They are silent by design, so the text on screen carries the whole explanation and they can be watched anywhere.
Why Data Access Decides the Answer
A stronger model does not fix a wrong join. What an agent needs from the data platform before it can be trusted.
- A better model will not fix a wrong join
- Ground the agent, do not coach it
- What data access has to deliver
The Semantic Layer
How a semantic layer turns raw lakehouse tables into business meaning an agent can rely on.
- Raw tables carry no meaning
- Resolve meaning before the query runs
- The same question, the same number
Governance for Agents
Why prompt rules are not access control, and how policy enforced at query time bounds what an agent can see.
- A prompt rule is not a permission
- Filter at query time, not after
- The agent borrows the user's rights
MCP for Data
What the Model Context Protocol standardizes, and what its primitives look like when the system on the other end is a lakehouse.
- Every agent wrote its own connector
- A server exposes a described surface
- The tools are data operations
RAG or a Semantic Layer?
Where retrieval-augmented generation is the right tool, where a semantic layer is, and why most real systems use both.
- Passages come back, totals do not
- A number has to be computed
- Route the question to the right tool
Snapshots Make Agent Work Auditable
How Iceberg snapshots let you reproduce exactly what an agent saw, months after it said it.
- The table moved on without you
- Iceberg keeps every committed state
- Log the snapshot with the answer
Letting Agents Write Safely
The write-audit-publish pattern on Iceberg branches, and where the human gate belongs.
- Reads are cheap to get wrong
- Write on a branch, publish on pass
- Keep a human on the publish step