Developing
A shared context layer is the idea: agents answer better when they know what the business terms mean.
In 60 seconds
- On June 2, 2026 Snowflake described an "agentic enterprise" with CoWork, a personal agent for knowledge workers, and CoCo, a coding agent.
- Cortex Sense is a shared context layer; Snowflake's own tests, per one summary, raised retrieval accuracy from 47% to 83%.
- Iceberg v3 became generally available with Snowflake-managed storage; Datastream is a Kafka-compatible streaming service.
- Security features include Agent Identity and AI Security Posture Management.
Impact Layer
- What changed
- Snowflake framed its platform around agents and the data context they need.
- Why it matters
- Business definitions held in one place help marketing and finance get consistent answers from agents.
- Who is affected
- Data and analytics teams on Snowflake.
- Martech stack area
- Marketing Automation, AI
- Stack impact
- Write down your key business definitions; a context layer is only as good as they are.
- Business impact
- Moderate Accuracy numbers are Snowflake's own tests as relayed by a summary.
- Related technologies
- Related vendors
- What to watch
- Customer results for Cortex Sense and adoption of Iceberg v3.
Snowflake Summit 2026 took place in San Francisco in early June. SiliconANGLE’s report is dated June 2.
What was announced
- CoWork, renamed from Snowflake Intelligence: a personal AI agent for knowledge workers with Deep Research, Artifacts and User Memory.
- CoCo, formerly Cortex Code: a coding agent now supporting desktop, mobile, Slack, Excel, VS Code and Claude Code.
- Cortex Sense: a shared context layer giving agents access to business definitions and enterprise context.
- Iceberg v3 support, generally available, with Snowflake-managed storage and an enhanced Horizon Catalog.
- Security: Agent Identity, AI Security Posture Management and centralized governance to prevent data exfiltration.
- Datastream, a managed Kafka-compatible streaming service, and Cortex Training, managed GPU infrastructure for fine-tuning open-source models.
Christian Kleinerman, SVP of product: “Models keep changing, and capabilities keep advancing, but the data is constant.” Another summary says Snowflake’s tests show Cortex Sense raising agent retrieval accuracy from 47% to 83%; that figure is Snowflake’s and we did not read its source.
Our read
The practical idea is the context layer. When two agents define “active customer” differently they give different answers. A shared definition store fixes that, and it only works if your team writes definitions down. For marketing teams on Snowflake, start by listing the ten terms (customer, active, churned, qualified lead) that reports disagree on.
Sources
- SiliconANGLE: AI agents, open data and governance take center stage at Snowflake Summit (opens in a new tab)
Source of the product list and the quotation by Christian Kleinerman.
- Atlan: Snowflake Summit 2026, all the announcements and what they mean (opens in a new tab)
Cited in the search results for the 47% to 83% accuracy figure. Not read directly.