Snowflake Summit 26 in San Francisco made Snowflake’s AI strategy much clearer. The company is trying to make enterprise data, governance, agents, and developer workflows operate inside one trusted execution environment.
The June event brought Snowflake CEO Sridhar Ramaswamy together with Anthropic President Daniela Amodei and other enterprise leaders, but the main story was not only model access. Snowflake used the event to argue that agentic AI will succeed only when agents can work from governed data, verified identity, and live operational streams.
Key Announcements
CoWork and CoCo Put Agents Into Daily Work
Snowflake used Summit to push two agent surfaces forward. Snowflake CoWork, formerly Snowflake Intelligence, is now positioned as a personal work agent for business users, while Snowflake CoCo, formerly Cortex Code, is Snowflake’s AI coding agent for developers and technical teams.
The distinction is real. CoWork is aimed at people who need to ask questions, generate artifacts, and interact with governed data without learning SQL. CoCo is built for teams creating data applications, automations, and AI workflows on Snowflake. CRN reported that Ramaswamy described faster implementation and time-to-value as part of the opportunity, with partners playing a central role in turning these capabilities into customer outcomes.
Snowflake also expanded where these agents can work, including plugins and integrations across tools such as VS Code, Claude Code, Slack, and Excel. That turns the data cloud from a destination into an agent surface that follows users into their existing workflows.
Horizon Context and Iceberg Target the Data Fragmentation Problem
Snowflake’s interoperability announcements focused on a familiar enterprise problem. AI agents are only as useful as the data and context they can trust. Snowflake announced general availability support for Apache Iceberg v3 and Snowflake Storage for Apache Iceberg Tables, along with expanded Horizon Catalog capabilities powered by Apache Polaris.
Christian Kleinerman, Snowflake’s EVP of Product, framed the issue directly in Snowflake’s announcement, saying that data fragmentation becomes the constraint as AI adoption accelerates. The company’s answer is a governed, logical data copy that can be accessed across Snowflake, external lakes, and open systems without forcing constant data movement.
Horizon Context extends that argument into semantics. Snowflake is trying to collect metadata, lineage, and business definitions so AI agents can reason from shared enterprise meaning rather than isolated tables. That semantic layer is the practical foundation behind the agentic enterprise language.
Datastream, Openflow, and Cortex Training Move Toward Live AI Operations
Snowflake also pushed deeper into the operational layers that feed AI systems. Datastream, a managed Kafka-compatible streaming service, is designed to bring real-time data into Snowflake without requiring teams to run separate Kafka infrastructure. Openflow, built on Apache NiFi, addresses batch and streaming integration across enterprise systems.
AI agents need more than historical warehouse data. They need timely signals from applications, events, and operational systems. Datastream and Openflow give Snowflake a stronger claim that it can support live data movement and ingestion inside its own governance model.
Cortex Training added another layer. Snowflake is offering managed infrastructure for fine-tuning open-weight models such as Qwen and Mistral using proprietary enterprise data inside Snowflake. For technology leaders, the relevant question is whether model customization can happen without creating a separate GPU estate and a separate governance problem.
Security Moves From User Identity to Agent Identity
Snowflake also used Summit to put agent security closer to the center of the platform. Announcements around AI Agent Identity, AI Security Posture Management, data exfiltration protections, and ransomware protection point to a platform that treats agents as actors that need their own rules.
Agent identity is the most important concept here. Traditional access models were built around human users and service accounts. Agentic systems create a different problem because agents may query data, call tools, and trigger downstream actions. Snowflake’s approach gives agents distinct identity and permissions rather than simply inheriting a user’s full access.
That move is necessary if CoWork, CoCo, and customer-built agents are going to handle real enterprise work. Without agent-specific identity, audit trails and access boundaries will be too blunt for production use.
Strategic Insights
Snowflake Is Selling Governed Context, Not Just AI Access
The strongest thread across Summit was Snowflake’s insistence that model choice is not the center of enterprise AI. The company is positioning itself around the data, metadata, semantics, and identity controls that determine whether models can be used safely inside business workflows.
The announcements fit together for that reason. CoWork and CoCo create agent surfaces, Horizon Context and Iceberg make data and metadata more usable, and Datastream and Openflow feed those systems with live signals. Agent Identity and Trust Center capabilities define who or what can act. Snowflake is building toward a control layer for enterprise intelligence, not simply adding AI features to a data warehouse.
Openness Is Becoming Part of the Platform Pitch
Snowflake also leaned hard into interoperability. Iceberg v3, Horizon Catalog, Polaris, and MCP-related ecosystem moves all speak to customers that do not want AI strategies trapped inside one proprietary data path.
The tension is that Snowflake still benefits when more data, workflows, and agent interactions stay close to its platform. The company’s answer is to make openness part of the architecture while keeping governance, context, and agent execution centered on Snowflake. That is a stronger pitch than pure lock-in, but it still asks enterprises to decide how much operating control they want one vendor to hold.
The Undercurrent
The operating tension at Snowflake Summit 26 was openness versus control. Snowflake wants to make data more interoperable across clouds, tools, and external systems, while also making Snowflake the place where governance, agent identity, and AI execution come together.
That balance will determine adoption. Enterprises want open data access and model choice, but they also need consistent policy, lineage, and auditability. Snowflake’s bet is that customers will accept more platform gravity if it reduces the risk of fragmented AI deployments.
Why It Matters
For technology leaders, Summit 26 made the AI roadmap more concrete. The work is no longer limited to choosing models or experimenting with copilots. Teams need to prepare data estates so agents can understand business context, access live signals, and leave a usable audit trail.
Snowflake’s direction also raises a governance ownership question. Data teams, security teams, and application builders will need a shared model for agent permissions, semantic definitions, and operational responsibility. Without that, agent adoption will recreate the same fragmentation Snowflake is trying to remove.
What’s Next
Start by identifying one workflow where governed data access and agent assistance could reduce real friction, such as financial analysis, customer segmentation, or data engineering automation. Use that pilot to test whether Snowflake’s context, identity, and governance controls are strong enough for production use.
Enterprise context is the constraint. Agents will fail when business definitions are inconsistent, metadata is thin, or access rules are inherited too broadly. Snowflake’s announcements give customers more tools, but the harder work is making the organization’s data understandable enough for agents to act on it safely.
