Creating a Data Warehouse That AI Agents Love
What would a warehouse look like if its most demanding users were AI agents? Jordan Tigani has a hypothesis. He's co-founder and CEO of MotherDuck, a founding engineer on BigQuery, a SingleStore alum, and now building a DuckDB-based cloud warehouse that can split work between local and cloud execution. Bryan and Jordan get into agent-first database design, scoped context for LLMs, MCP, and why analytics may be moving into the chat window.
In our conversation, we discuss:
- The local-to-cloud DuckDB loop: prototype on small samples, then productionize just by changing the database name
- Why Jordan moved from NL-to-SQL skeptic to agent-loop maxi once models stopped one-shotting SQL
- Why remarkable agent loops still create a low-trust product problem
- Agent tenancy: one cloud DuckDB instance per user or agent, from tiny pulse boxes to Snowflake-3XL-sized machines
- Workspace scoping: attaching only relevant data instead of drowning the LLM in a million databases
- Why tables still matter when Claude can stare at the lake: quality, trust, and performance
- Claude as the analytics surface: remote MCP servers, Mermaid schema diagrams, and Slack, Notion, Linear context
Three takeaways from this conversation:
1. The local loop is the wedge. MotherDuck tries to preserve the laptop feel as execution moves into the cloud, so local prototypes can become production workloads by changing the database name. The product move is simple: keep the developer’s loop small while the compute moves somewhere larger.
2. Agents turn SQL generation into investigation. Jordan’s old objection to NL-to-SQL was context: a one-shot model had to infer table relationships and analyst conventions in a single pass. The agent version looks more like onboarding a new analyst who reads docs, tests joins, fails, and tries again.
3. Scope is a product primitive for agents. MotherDuck’s workspace model makes the LLM see only attached data, the data chosen for that workflow, rather than the whole org universe of nameable objects. Jordan links that smaller search space to useful answers with far less semantic scaffolding.
