Agents over Data: Why SQL Is Not Enough
Агент над данными: почему одного SQL недостаточно
Episode participants
A solo episode without invited guests.
What we will explore
A scheduled Research Insights Made Simple episode walks through the longread “Agents over data: why SQL is only the beginning”. It opens with a teaching example: the query runs without errors, yet joining orders to refunds yields revenue of 350 instead of 250.
We then follow the functions of a harness over a data platform: context discovery, metric definitions, execution, verification, authority and observability. Next come the Snowflake and Databricks comparisons and why a shared model does not make tools and metadata access identical.
The central question is how to compare harnesses on one model: which benchmarks to take first, what to record in a run passport and how to count cost per accepted result. The experiment protocol is a method, not a report of the author's own runs.