Composable Semantic Layers for Analytical Agents

This article argues that AI analytical agents require more than a simple metric catalog; they need composable business logic that handles cohorts, time windows, exclusions, and period comparisons. It contrasts traditional metric catalogs with composable semantic layers, citing tools like dbt, Cube, and Dremio. The piece details architecture patterns, common failure modes, guardrails for agentic use, and operational checklists for engineers, data owners, and executives. It emphasizes that semantic contracts must be explicit enough for both humans and software, enabling agents to combine reusable operations without rewriting logic. The content is technical, focused on data engineering, AI platforms, and production analytics.