AI Changes the Data Interface, Not the Query Engine
InfoQ, Thursday, October 22nd, 2026
A disaster-recovery analysis project shows agents still need deterministic query engines, now embedded in the agent workflow.
The authors built an agent to analyze a disaster-recovery islanding exercise, storing 30-second metrics as Parquet in object storage, using DuckDB for aggregation and letting the agent classify failover outcomes.
This raised the question of whether lakehouses and SQL are still needed.
Their answer is that the query engine did not disappear but became an embedded component: agent-generated read-only queries are checked against an allowlist, run in a sandbox, and turned into structured fact packets for the model to reason over.
They separate deterministic computation from probabilistic reasoning and offer four signals for when an embedded engine is enough and when a managed lakehouse is justified.