What Is Snowflake Cortex Sense, and Where Does It Fit in Your Stack?

Cortex Sense is Snowflake’s version of context management.

Cortex Sense learns business meaning from signals Snowflake already has, its query history, metadata, and dashboards, so agents get usable context on day one rather than after a modeling project. This replaces the process of building a Snowflake Semantic View. Semantic Views can be a good way of creating semantic understanding for agents, but they require a great deal of manual setup, can be difficult to manage, and start going stale immediately.

What is Snowflake Cortex Sense?

When a business user asks CoWork a question or a developer asks CoCo to build something, Cortex Sense supplies the definitions and relationships the Cortex agents underneath need to answer correctly.

Two Snowflake experiences draw on Cortex Sense:

1. CoWork, formerly Snowflake Intelligence, is the interface for knowledge workers. It gives business users access to Cortex agents like Cortex Analyst and Cortex Search, which do the reasoning behind an answer.
2. CoCo, formerly Cortex Code, is the development environment for engineers, closer to an IDE than an agent. It calls the same Cortex agents to generate pipelines, SQL, and dbt models.

Both were announced alongside Cortex Sense. The agents working underneath each surface read from the same context layer at runtime.

The big selling point is automatic context with no manual authoring. According to Snowflake, Cortex Sense learns how a business defines its data, including concepts like revenue definitions, fiscal calendars, and snapshot tables, and makes that available to agents from day one.

Cortex Sense vs Horizon Context: what’s the difference?

Cortex Sense and Horizon Context are easily conflated because both are described as context. 

However, they do different jobs:

• Horizon Context is the governed store. It extends Snowflake’s Horizon Catalog to hold business definitions, metrics, lineage, and glossary terms, and it does reach beyond Snowflake, ingesting metadata from a limited set of outside sources such as dbt, SAP, and a handful of BI tools.
• Cortex Sense is the runtime layer that reads that store, combines it with the other signals listed in the next section, and serves the right context to an agent when a question arrives.

	Horizon Context	Cortex Sense
What it is	A governed store of business definitions, metrics, lineage, and glossary terms, extending Horizon Catalog to a limited set of non-Snowflake sources	A runtime layer that reads that store plus query history, metadata, and BI definitions
Its role	Holds and governs the definitions	Retrieves and serves the fitting definitions to agents at query time
Who reads it	BI tools, applications, and Cortex Sense	The Cortex agents behind CoWork and CoCo

One way to think about the distinction: Horizon Context is the library, and Cortex Sense is the librarian who fetches the right book when asked. The reach matters. Horizon Context covers a defined set of connected sources, which is a narrower footprint than a dedicated context platform maintains across an enterprise stack. Both are in preview as of mid-2026.

How does Cortex Sense work?

Snowflake describes Cortex Sense as building its context from signals already present in and around Snowflake, then serving the relevant pieces to an agent at the moment a question arrives. It is designed to draw on four main sources:

1. Query history: The patterns in how analysts filter, join, and aggregate reveal business definitions implicitly. If teams consistently filter revenue by a particular date field, Cortex Sense infers that field as a key anchor.
2. Object metadata: Table names, column names, descriptions, and relationships registered in Snowflake’s schema.
3. BI dashboard definitions: Metrics and measures from tools like Power BI and Tableau, the closest proxy for business-validated definitions already in Snowflake’s orbit.
4. Horizon Context semantic views: Governed definitions already registered in Horizon Context, including the metadata it has pulled from outside Snowflake.

Snowflake has also described a self-correcting loop. When signals conflict (for example, when different teams calculate daily active users differently), Cortex Sense surfaces the conflict and asks a human to settle it, then updates its understanding. To speed adoption, Snowflake says it will ship with prebuilt plugins for domains like finance and sales that bundle skills, business logic, and connectors.

How Cortex Sense, CoWork, and CoCo fit together

CoWork and CoCo are the surfaces people interact with. Cortex Sense is the shared substrate both of them read. One context layer sits underneath, and two agents sit on top of it.

Rather than building context separately for each agent, Snowflake assembles it once in Cortex Sense and lets both CoWork and CoCo consume it. The upside is consistency inside Snowflake. The limitation is that the same design keeps that context tied to Snowflake’s agents.

Benefits: What does Snowflake say Cortex Sense delivers?

Snowflake frames Cortex Sense around one headline outcome, accuracy, plus a set of benefits that follow from removing manual setup. Because it is in private preview, these are Snowflake’s own figures and claims rather than independently verified results.

Accuracy is the central claim. In internal testing on complex enterprise queries, Snowflake reports that CoWork and CoCo answered correctly 47% of the time without Cortex Sense and 83% with it. SiliconANGLE reported the same story with different framing, roughly 24% for a generic coding agent alone and about 86% with Cortex Sense out of the box. 

The numbers vary, but the direction is consistent: context determines agent quality. The closer an agent sits to governed business context, the more often it answers correctly.

The rest of what Snowflake attributes to Cortex Sense follows from the automation:

• Context from day one: By learning meaning from query history and metadata instead of a hand-built semantic model, Cortex Sense is meant to give agents usable context immediately, which Snowflake positions as solving the cold-start problem that stalls most agent rollouts.
• Consistency across agents: Because CoWork and CoCo read from the same layer, they work from the same definitions rather than each interpreting a metric differently.
• Less manual upkeep: The self-correcting loop and continuous signal mining are meant to keep context current as the business changes, rather than leaving it to go stale after a one-time documentation push.
• Faster path to production: Prebuilt plugins for domains like finance and sales bundle skills, business logic, and connectors, which Snowflake says can take teams from concept to a working agent in days rather than weeks.

These benefits describe a capability that’s still in private preview, measured on Snowflake’s own data. Standard CoWork and CoCo deployments run at the lower accuracy baseline until Cortex Sense is enabled.

The catch: Cortex Sense only serves Snowflake’s agents

The context Cortex Sense builds is consumed inside the Snowflake ecosystem. It can draw signals from some sources outside Snowflake, but the context it produces does not travel outward. An agent running elsewhere in your stack, whether that is Databricks Genie, a custom-built agent, or an assistant your team uses outside Snowflake, cannot call that context.

Model choice is a separate question from context portability. Snowflake lets teams point CoWork and CoCo at models other than its own, with some limitations. That changes which model does the reasoning inside the Snowflake surface. It does not make the context portable, and Snowflake has not documented whether Cortex Sense will serve context to those alternative models the way it does to Cortex agents.

Cortex Sense lives with the agent infrastructure, not in a shared metadata layer that any tool can call. That makes it additive for Snowflake work and genuinely useful there. It also makes it non-portable. The context you build for Snowflake’s agents stays with Snowflake’s agents.

Every platform building its own context layer—it creates a new need

Cortex Sense is not an isolated move. It is one instance of a pattern playing out across the industry. Snowflake has Horizon Context and Cortex Sense. Databricks has Unity Catalog and Genie Ontology. Microsoft is building IQ. Salesforce has Data Cloud. Each of these is a context layer, and each is built to serve its own vendor’s agents.

“Nearly every major platform vendor is investing in some version of a context layer.”
— Constellation Research

The shared realization is that enterprise AI needs more than access to data. It needs enterprise context, a machine-readable understanding of how the business operates. The complication is that when every platform builds its own version, context fragments across silos. 

Constellation Research notes that many customers continue to want a neutral layer that sits above individual platforms, and names DataHub among the platform-independent options they are asking for.

Why you still need a context platform across your whole stack

The platform-native context layers are additive, not portable. That is the case for a foundation beneath them. A context platform sits underneath every platform and serves one set of governed, validated context to every agent, regardless of where that agent runs.

This is the role DataHub is built for. It ingests context from across 100+ sources rather than one, maintains cross-platform lineage down to the column level, and enriches technical metadata with business definitions, validated query patterns, and institutional knowledge from tools like Confluence and Notion. Domain experts review and correct that context, and it is served to any agent through Model Context Protocol (MCP), GraphQL, and APIs and SDKs.

For a Snowflake customer, the two are complementary. Cortex Sense makes Snowflake’s agents sharper on Snowflake data. DataHub gives Snowflake Cortex the context that lives outside Snowflake and delivers the same governed context to non-Snowflake agents, so the whole estate works from one trusted source rather than a separate context silo per platform. 

Agents will keep changing. The context they rely on is worth building once, in a place every one of them can reach.

FAQs

Is Snowflake Cortex Sense generally available yet?

As of mid-2026, Cortex Sense is in private preview, not general availability. Snowflake’s CoWork and CoCo surfaces are available, but the Cortex Sense context layer beneath them is still rolling out through the preview program. Check Snowflake’s current documentation before assuming it is active in your account.

What is Snowflake CoWork?

Snowflake CoWork is Snowflake’s work interface for knowledge workers, formerly Snowflake Intelligence, announced at Snowflake Summit in June 2026. It gives business users access to Cortex agents that query governed Snowflake data in natural language, run multi-step research, publish dashboards, and take action across tools like Slack, Gmail, and Salesforce. Those agents read their business context from Cortex Sense.

What is Snowflake CoCo?

Snowflake CoCo is Snowflake’s development environment for engineers and data teams, formerly Cortex Code. It is closer to an IDE than an agent, calling Cortex agents to build Streamlit applications, data pipelines, SQL, and dbt models inside the Snowflake ecosystem. Like CoWork, it draws on Cortex Sense for the business context those agents need.

What’s the difference between Cortex Sense, Cortex Search, and Cortex Analyst?

They solve different problems:

• Cortex Analyst converts natural-language questions into SQL against structured data.
• Cortex Sense assembles business context for Snowflake’s agents
• Cortex Search is a managed search service for finding relevant text across your data, often used for retrieval-augmented generation (RAG)

In short, Cortex Sense is the context layer, and the other two are Cortex agents that draw on it to answer questions inside surfaces like CoWork and CoCo.

Can Cortex Sense deliver context to non-Snowflake agents?

No. Cortex Sense builds context for the Cortex agents behind CoWork and CoCo. Even though it can draw signals from some sources outside Snowflake, the context it produces stays within the Snowflake ecosystem. Agents running elsewhere in your stack, such as Databricks Genie or a custom-built agent, cannot consume it directly. Note that this is about where context can travel, not which model you use. You can point CoWork and CoCo at non-Snowflake models, but that does not make the context available outside Snowflake.

Do you still need a context platform if you have Cortex Sense?

For most enterprises, yes. Cortex Sense makes Snowflake’s own agents more accurate, but most enterprise data estates run agents and tools well beyond Snowflake. A context platform like DataHub sits underneath every platform-native context layer and serves one set of governed, validated context to every agent, regardless of where it runs. It ingests across 100+ sources, maintains cross-platform lineage, and delivers context through Model Context Protocol (MCP). Cortex Sense and a cross-platform foundation are additive, not competing.

How does DataHub work alongside Snowflake Cortex Sense?

DataHub acts as the cross-platform context foundation beneath Snowflake and every other platform in your stack. It gives Snowflake Cortex the broader context that lives outside Snowflake, including business definitions, validated query patterns, full technical lineage, and unstructured documentation from tools Snowflake does not touch, capturing deep process knowledge that never lands in a schema. That context is validated by domain experts and served to any agent through MCP, so Snowflake-native and non-Snowflake agents work from the same trusted context.