A traditional semantic layer in BI (like LookML or dbt) handles specific metric definitions — revenue, churn, ARPU — using hand-constructed syntax connected to a single BI tool. A modern context layer must become a superset of this: canonical entities, identity resolution, specific instructions to dissect tribal knowledge, and governance guidance. This is Structure plus reasoning beats flat similarity for complex domains applied to enterprise data — flat metric definitions aren’t enough when agents need to reason across disparate sources.
The context layer becomes a multi-dimensional corpus where code lives alongside natural language. Just as developers set up .cursorrules files to guide coding agents, data practitioners maintain rules and guidelines for data agents. This mirrors the pattern in Skill graphs enable progressive disclosure for complex domains — layered, structured context beats monolithic definitions. The context layer can be exposed to agents via API or MCP, connecting to Domain-specific skill libraries are the real agent moat, not core infrastructure — the structured knowledge about a specific enterprise’s data is the moat, not the agent infrastructure.