The modern data stack is the infrastructure layer that made enterprise data accessible. Ingestion, transformation, warehousing, and visualization stitched together. It solved the plumbing problem. The core stack: a cloud data warehouse ([[Snowflake]], BigQuery, Redshift) as the storage and compute layer, [[dbt]] for transformation (SQL-based, version-controlled, testable), an ELT tool (Fivetran, Airbyte) pulling data from source systems, and a BI layer (Looker, Tableau, Metabase) on top. That combination became the de facto enterprise data architecture between roughly 2016 and 2023. The thesis was clean data = self-serve analytics. Put everything in one place, model it properly, let analysts write SQL and build dashboards. It worked, partially. Data was more accessible than ever. But the promise of "anyone can answer any question" didn't land. Analysts became bottlenecks. Dashboards went stale. Tribal knowledge never made it into the warehouse. The [[semantic layers]] were supposed to close the gap. Define metrics once (LookML, dbt metrics), expose them consistently across tools. But semantic layers are still static. They cover what's defined, not what's implied. They don't capture why certain tables are authoritative, what the fiscal calendar edge cases are, or how the business actually reasons about its data. When [[AI agents]] arrived, the modern data stack was the foundation they ran on and the first thing that broke them. An agent hitting a well-built warehouse still couldn't answer "what was revenue last quarter" reliably, because the answer requires business context that lives outside the schema. That gap is what the [[Context Layers MOC]] is about. The modern data stack era produced two durable assets: clean, centralized data and a generation of data engineers who understand transformation logic deeply. Both are inputs to building the [[context layers]] that agents actually need. --- #deeptech #systems #kp