Trustworthy AI in the Agentic Lakehouse: from Concurrency to Governance

Infrastructure design is foundational to AI trustworthiness: The paper argues that enterprises' hesitation to deploy AI agents on production data stems not primarily from AI capability limitations, but from inadequate underlying infrastructure. By building data systems specifically designed for agent access patterns, governance and trust can follow naturally rather than being bolted on afterward.

Traditional data systems create safety gaps for AI agents: Standard lakehouse architectures were not designed for how AI agents access and modify data. The paper demonstrates that directly applying existing database solutions (like Multi-Version Concurrency Control) fails in decoupled, multi-language environments where agents operate, leaving critical vulnerabilities in production systems.

Isolation and governance must be redesigned for agent workflows: Bauplan, the proposed agent-first system, reimplements data and compute isolation specifically for AI agent operations. This architectural approach enables enterprises to maintain the correctness guarantees and audit trails necessary for trustworthy AI—addressing a key barrier to production deployment.

Self-healing pipelines demonstrate practical trustworthiness: The paper's reference implementation shows that when infrastructure properly couples agent reasoning with built-in safety guarantees, systems can automatically detect and correct errors, reducing the need for constant human oversight while maintaining reliability.