Rethinking the Semantic Layer in an Agent-Native Data Stack
You don't need another semantic layer. All you need is code.
For decades, the semantic layer existed because humans could not read SQL. Semantic layer tools translated raw schemas into business language so analysts and executives could ask questions without engineering training. Agents are now becoming the primary consumer of data. They read code fluently, write SQL fluently, and compose queries on demand. The vendor response has been to rebrand the semantic layer as a "context layer" for agents, a surface that sits between governed metric definitions and the LLM. OSI, launched by Snowflake with 40+ partners in early 2026, formalized a vendor-neutral YAML format for this layer. Every major data vendor now claims a context layer position.
In this webinar, we will examine what context actually means when agents are the consumers, and whether a separate semantic layer earns its place in an agent-native stack. The session will walk through the current semantic layer landscape, the architectural split between headless and platform-native approaches, and the framing of the "context layer" as a superset that includes entities, identity, governance, and tribal knowledge. From there, it will introduce a contrarian position: most of what a semantic layer provides for agents already lives in well-written pipeline code. Schema contracts, lineage, computation logic, materialization policy, environment specifications, and business definitions can all be expressed as decorators, type hints, Pydantic classes, and docstrings.
Through real-world Bauplan pipelines, the session will show how a single Python file can carry the contracts, dependencies, and definitions that a semantic layer expresses across multiple YAML files and proprietary tools. It will discuss where the equivalence breaks: cross-tool portability for legacy BI consumers, synonym discovery, and aggregation correctness across joins. The discussion will close with a forward look at how the semantic layer collapses into a thin adapter for human-facing BI, while typed, branchable, executable code becomes the substrate agents actually operate on.