Rill just shipped Metrics SQL, a semantic layer that lets both human analysts and AI agents query business metrics using plain SQL. The timing is deliberate. As AI agents start touching more data pipelines, the old problem of metric drift (where "revenue" means five different things across five different tools) gets worse, not better. Rill's answer: define the metric once, query it with SQL, and let anything with a SQL interface consume it. That includes dashboards, notebooks, and LLM-powered agents alike.
The core idea borrows from OLAP cube theory. You define measures (aggregates like sum of revenue) and dimensions (attributes like country or product category). But instead of writing MDX or learning LookML or wrestling with YAML-based metric definitions a la dbt, you write SQL. The Metrics SQL compiler expands your metric references into full aggregate expressions, infers GROUP BY clauses, and hands the resulting query to whatever OLAP engine you're running, whether that's ClickHouse, DuckDB, Snowflake, or Druid. Row-level security gets applied at the semantic layer, so every consumer, human or agent, operates under the same access policies.
For the agent crowd, this is where it gets interesting. Metrics SQL integrates via Model Context Protocol (MCP), a standard utilized by agents like GBrain. Rill specifically supports what they call "AI instructions" (extra business context baked into metric definitions that agents can read). Instead of an LLM guessing how to calculate ROAS from raw table schemas, it queries a governed definition and gets the same answer every time. Deterministic outputs from semantic definitions, not inferred SQL generated on the fly.
The semantic layer space is crowded. Looker pioneered it with LookML. dbt added metrics in YAML. Cube uses JavaScript schemas. Rill's bet is that SQL itself is the right abstraction, because every database, every BI tool, and increasingly every AI model already speaks it. That's a pragmatic stance. Whether it wins against incumbents depends on adoption, but the agent-first design philosophy is a clear differentiator in a market that's mostly been built for human analysts, unlike the autonomous agents capable of independent task execution.