Snowflake EVP of Product Christian Kleinerman used Summit 26 to put a name on the failure mode every team running agents on top of a warehouse has hit this year: the same question routed through SQL, BI and an agent prompt returns three different numbers, all delivered with the same confident tone. The company's answer is to stop patching this at the agent and fix it at the catalog. Horizon Context holds what customers explicitly declare. Cortex Sense holds what the platform derives from data and query patterns. The two layers are deliberately kept separate because, as Mike Leone at Moor Insights put it, you cannot trust declared and derived context the same way, and treating them identically is how agent answers go quietly wrong in production.
The two-layer split, in plain terms
Horizon Context is the explicit side. It is built on the Select Star acquisition Snowflake closed earlier this year and pulls metadata from Postgres, SQL Server, Tableau and Power BI into the Horizon Catalog. The point is that every agent, BI tool and external system reasons against the same governed definition of revenue or active customer rather than improvising over physical schemas. Semantic View Autopilot sits on top and creates and refines semantic views automatically, which directly targets the historical problem that hand-authored semantic layers rot the moment a column gets renamed or a join changes.
Cortex Sense is the implicit half. It enriches context from customer data and usage on an ongoing basis with no manual authoring required, so the default experience improves before any curation work happens. Kleinerman's framing is that Horizon Context is everything explicit and declared by customers and Cortex Sense is anything implicit and derived by Snowflake. Reconciliation between the two is the part that has to be observable, because the moment an agent answer mixes a declared definition with a derived signal without an audit trail, trust collapses.
Why Snowflake put it in the catalog
The competitive map is dense. Microsoft is shipping Fabric IQ as a business ontology accessible via MCP, so any vendor's agent can call it. Redis is pushing Iris as a context and memory platform sitting between agents and data. Pinecone Nexus compiles enterprise data into task-specific artifacts before agents query. Each vendor has a different bet on where the context layer should live, and Snowflake's bet is that the catalog is the right home because governance, lineage and policy already sit there. Devin Pratt at IDC backed the architectural call, noting that enterprises do not need another silo of semantics but a layer that is governed, portable and trustworthy enough to audit. The portability commitment is the Open Semantic Interchange, which Kleinerman said Snowflake is leading the effort on so that Horizon Context does not become another lock-in tax sitting on top of the warehouse.
The demand curve behind the announcement
VentureBeat's VB Pulse Q1 2026 data showed hybrid retrieval intent tripling from 10.3% in January to 33.3% in March, the fastest-growing strategic position in the dataset. That curve is the demand signal driving every catalog and database vendor toward a context layer this quarter. Two years of retrieval infrastructure produced faster vector search but no shared definition of what data means, which is why confident wrong answers keep surfacing in production reviews. The Summit 26 release also included Data Stream as a Kafka-compatible managed streaming service, expanded Apache Iceberg interoperability, and updates to the CoCo and Cowork agent products that consume context through Cortex Search as a tool call.
How we are piloting this in Q1 2026
Our plan is concrete. We are scoping a Horizon Context pilot against a single high-value domain where agent answers are already inconsistent, and we are picking revenue recognition because variance there is visible to finance within a week. The success metric is a measurable drop in answer variance across three surfaces: the BI dashboard, a direct SQL query and a Cortex agent. If variance falls below five percent across thirty representative questions, the pilot graduates and the next two domains queue up. Portability is a hard requirement, not a preference. We are also wiring the same domain into a second catalog through the Open Semantic Interchange path so that we can prove definitions move without translation losses. Budget envelope is roughly 180 thousand dollars across six weeks, covering Select Star integration consulting, internal data engineering hours and a part-time analytics engineer to own the semantic views. The decision gate is February 15. If reconciliation between Horizon Context and Cortex Sense is not observable by then, meaning we cannot trace an agent answer to its layer of origin, we hold the rollout and revisit at Snowflake's next product update rather than expanding the pilot on faith.
Pricing, portability and the dbt question
There is a vendor-management story under all of this. If the context layer becomes the de facto governance plane for agents, budget consolidates. Spend that previously went to standalone catalog vendors, semantic-layer startups and BI governance tools gets pulled toward whichever warehouse holds the authoritative surface. Teams that already standardized on dbt for transformation face a real choice between dbt's semantic layer and Snowflake's Horizon Context, and the answer turns on whether the organization is single-warehouse or multi-warehouse. Multi-warehouse shops will want dbt to stay authoritative and Horizon to mirror. Single-warehouse Snowflake shops will probably let Horizon win because the catalog is already where lineage and policy live. Either way, Open Semantic Interchange is the contract that keeps the choice reversible, and the test of whether Snowflake means it is whether competing catalogs can read and write Horizon definitions without losing fidelity.
What has to be true by June 2026
Three conditions decide whether this becomes load-bearing infrastructure or a slide. Semantic View Autopilot has to produce views that survive real query loads, not demo data. Reconciliation between declared and derived context has to be observable to an auditor, which is the bar Pratt set when he said the win condition is tracing every answer back to its layer of origin. And Open Semantic Interchange has to ship with at least two non-Snowflake catalogs in production by June 2026, otherwise portability is a press release. Miss any of the three and the catalog-as-context-layer thesis stalls, agent answers stay inconsistent, and the budget that consolidated toward Snowflake fragments back out to point tools within two quarters.
