Autonomous LLM Agents Over Iceberg Lakehouses: Governance and Cost Guardrails

Apache Iceberg has become a widely used open table format in cloud data platforms, offering ACID transactions, schema evolution, hidden partitioning, and time-travel snapshots over data stored in Google Cloud Storage. Connecting autonomous LLM agents to such lakehouses introduces risks, including runaway queries that can incur large scanning costs, hallucinated partition columns, and destructive DDL or DML statements. To address these, the article proposes a production-oriented agent architecture built on Google's Agent Development Kit patterns, Gemini 2.5 Flash, BigLake Iceberg REST Catalogs, and AST-based guardrails. The agent never directly touches data files or runs raw SQL against storage, instead working within a deterministic, three-tier execution loop that includes discovery, validation, and execution stages.