CoCo and CoWork are Snowflake's answer to a cost most organizations have never actually measured: the time a question spends waiting in a queue before someone with the right access and skills can answer it. At Summit 26, Snowflake made the case that this queue can be closed by putting agents inside the governed data platform itself, rather than bolting them on near it.
That queue is easy to miss because it never shows up in a process diagram, but it does show up everywhere once you actually start looking for it. A regional manager wants to know why margins slipped in one territory, field rep needs a briefing before a customer meeting, and a planner needs last quarter's actuals reconciled against forecast. None of these are difficult questions, but all of them have to travel, from the person who has the question to the person who can answer it, and back again. The analysis itself might take twenty minutes. The round trip takes four days. That round trip is the cost; not the query, the queue.
CoCo and CoWork are built to close that gap. This article covers what they actually do, what changed when Snowflake renamed them at Summit 26, where the business value shows up in practice, and the five things that determine whether any of it works inside your environment rather than just in a demo.
What Changed at Snowflake Summit ‘26
Both Cortex Code and Snowflake Intelligence were renamed on June 2, 2026. Cortex Code became CoCo. Snowflake Intelligence became CoWork. However, one of those renames is just cosmetic. Snowflake has confirmed that CoCo’s functionality and architecture are unchanged; most practitioners were already calling it CoCo, and the company made the nickname official.
The other renaming goes beyond a name change and reflects a broader repositioning and expansion in capability. Snowflake Intelligence was one component inside the Cortex AI suite: a conversational interface for asking questions of governed data. CoWork is a top-level product positioned alongside CoCo, and the capability set moved with the name. It now supports multi-step investigations, reusable artifacts, and actions across connected systems, with recurring automations currently available in Public Preview.
TL; DR
Cortex Code --> CoCo
No changes in functionality or architecture
Snowflake Intelligence --> CoWork
Snowflake Intelligence was a way to ask; CoWork is a way to act
This matters practically. If your internal documentation, enablement material, or architecture diagrams still say Snowflake Intelligence, they are describing an older and a considerably smaller, less capable product.
What is Snowflake CoCo and CoWork?
The core principle here is that one agent builds governed capability, and the other puts it directly in the hands of the business. Snowflake describes the pair as an agentic control plane. In business terms, the relationship is simpler: CoCo creates governed capability inside Snowflake's operating context, and CoWork puts that capability in the hands of people who previously waited in a queue for it.
CoCo builds. CoCo is the builder's agent, and it works where builders already work, in the Snowsight browser interface, in a native desktop application for macOS and Windows, and in a command-line interface. It also exposes an agent SDK and supports the Model Context Protocol and Agent Client Protocol, so other tools can drive it.
The distinction is not simply that CoCo can write SQL. CoCo operates with Snowflake-specific context, including roles, privileges, schemas and platform conventions, and can draw on additional environment context such as lineage where relevant. A general-purpose coding assistant may understand the repository without understanding that operating context, producing technically plausible code that still requires significant validation before it belongs in production. CoCo reduces that distance between generated code and something that can operate inside the governed environment. That is the difference between a suggestion and a change you can ship.
CoWork puts it to work. CoWork is the conversational surface the business uses. Underneath it, the Cortex Agent API routes a question to an orchestrator, which interprets intent and selects tools, Cortex Analyst for structured data, Cortex Search for unstructured content, and custom functions or procedures where an action is required.
There are three key CoWork capabilities that are worth naming from a business perspective:
Deep Research: For open-ended questions, CoWork breaks the question into sub-investigations, runs them in parallel across structured and unstructured data, and returns a cited report. Investigations can take up to ten minutes, which is the point. It is built for the questions that used to take days
Artifacts and automations: A chart or table can be saved, shared, and revisited without regenerating it. Recurring automations can also re-run work against fresh data and are currently available in Public Preview.
Action outside Snowflake: Through MCP connectors, CoWork reaches tools including Google Drive, Salesforce, and Slack. This is what moves it from advisory to operational.
Throughout, every answer traces back to the source data and the query that produced it, and existing row-access and column-level security policies apply automatically. Existing Snowflake data-access controls remain part of the execution path; the important work is ensuring those controls and semantic definitions are ready for agentic consumption.
How CoCo and CoWork Work Together
The loop closes because both halves read the same context. CoCo can help build and maintain governed data and semantic assets; CoWork can consume that context alongside Cortex Analyst, Cortex Search, and other tools to answer questions and execute workflows. When the two drift apart, so do the answers.
The value compounds across the chain: faster building matters, but the larger opportunity is removing the handoffs between question, insight, and action. Developer productivity is difficult to defend in a budget conversation because it is hard to attribute and easy to dispute. Cycle time is neither. If a monthly close, renewal review, or pre-call briefing measurably shortens, that number belongs to the business and survives scrutiny.
CoCo and CoWork in Action: From an Existing Power BI Model to a Governed Agent
The most instructive worked example is not a greenfield build. Rather, it starts from an asset most enterprises already own.
The scenario: A team had a mature Power BI semantic model, tables, active relationships, dozens of DAX measures, and wanted their sales organization to be able to interrogate it conversationally. The obvious objection was cost: rebuilding that model natively in Snowflake looked like months of work they had only just finished doing once.
Export the existing model: Working in CoCo's desktop application, the team exported the Power BI template file so CoCo could reverse-engineer it.
Let CoCo convert it: CoCo extracted the tables, relationships, and measures, generated a Snowflake semantic view, and flagged the objects it could not convert because they originated in a source outside Snowflake.
Stand up the agent: CoCo created a Cortex Agent pointed at the new semantic view, with seeded starter questions so the sales team had somewhere to begin.
Put it to work: The agent was opened in CoWork, and the sales organization began asking natural-language questions against it.
Elapsed time from export to a working agent: roughly one hour. The majority of that hour was spent reviewing and adjusting what the agent produced, not waiting for it to produce it.
The work did not disappear; it just changed shape, from construction to review. That change of shape is the honest version of the productivity story, and it is also why the review step is not optional. An agent that can build a semantic view in minutes can build a wrong one just as quickly.
Key Advantages of CoCo and CoWork
Governed within Snowflake: CoWork and the underlying Cortex services operate within Snowflake’s existing access-control model, including applicable RBAC, row-access, and column-level security policies.
Traceable answers: Every CoWork response traces back to the source data and the query that produced it, so the reasoning is auditable, not just the output.
Reusable, not disposable: CoWork artifacts can be saved, shared, and converted into recurring automations rather than regenerated from scratch each time.
Faster iteration on governed assets: CoCo can reverse-engineer and migrate existing semantic assets, like a Power BI model, rather than requiring teams to rebuild from zero.
Model flexibility within a governed environment: Snowflake can use different supported models and AI services while keeping access to enterprise data governed through the Snowflake platform.
Considerations & Limitations of CoCo and CoWork
Five things determine whether the possibilities and advantages are realized in your organization:
The default role trap: CoWork initializes with the user's DEFAULT_ROLE, not the role they happen to be using in Snowsight. If that default role has not actually been granted, Snowflake falls back to PUBLIC silently. Shared artifacts compound this: an artifact re-runs under the recipient's credentials, so two people can open the same object and see legitimately different numbers. This is defensible architecture and it is the correct behavior. It also means default roles must be audited before rollout, not after the first confusing meeting.
Context readiness: The agent inherits your definitions. If revenue is defined three different ways across three teams, the agent will confidently pick one. Semantic modeling is not preparation for the agent, it is the product the agent consumes, and it is where most of the real effort sits.
Cost becomes a consumption line: CoCo is billed on token consumption across every surface, including Snowsight, which was free earlier in 2026. This is a change worth flagging to anyone budgeting from guidance written before Summit. Consumption generally grows with workload complexity and adoption, making cost governance an important part of a successful rollout.
Access is broader than most teams assume: The COPILOT_USER and CORTEX_USER database roles are granted to PUBLIC by default, so everyone in the account has access unless someone decides otherwise. CORTEX_AGENT_USER is not, and must be granted explicitly. There is a trap in the obvious remedy: revoking CORTEX_USER from PUBLIC also removes access to Cortex AI functions such as AI_COMPLETE and AI_CLASSIFY. Tightening one door closes another. Per-user credit limits are the finer instrument.
Availability differs by surface: This is frequently missed. CoCo in Snowsight is generally available to commercial accounts and to FedRAMP Moderate and High, DoD, and KSA sovereign accounts. CoCo Desktop and the CLI remain limited to commercial accounts, excluding government, VPS, and sovereign deployments. For public sector organizations, the browser surface is reachable and the local surfaces are not, which shapes what a pilot can look like. All surfaces require cross-region inference to be enabled, so data residency belongs in the conversation early rather than late.
Pricing and Cost Considerations
CoCo and CoWork follow Snowflake's consumption-based AI pricing model rather than traditional per-seat licensing. Snowflake now prices applicable AI services using AI Credits, with consumption varying by the service, model and workload. Organizations should therefore treat adoption and cost governance as part of the same rollout.
The most important change to plan around: CoCo is now billed on every surface, including Snowsight, which was free earlier in 2026. Teams budgeting from pre-Summit guidance will underestimate cost.
Organizations should monitor consumption the same way they would any other Cortex AI service, tracking credit usage regularly and setting per-user limits rather than relying on broad access revocations, which can unintentionally remove access to unrelated AI functions like AI_COMPLETE and AI_CLASSIFY.
What Comes Next
The loop does not stop at Snowflake's own agents. The same governed context that makes CoCo and CoWork trustworthy is what allows other systems, planning, CRM, service management, to participate in it rather than sit beside it, each with its own credentials, its own audit trail, and its own bill. That raises the more interesting question, and the one enterprises will be answering next: not what a single agent can do, but what an estate of them can be trusted to do together.
To get started at your organization, you’ll want to find the first use case. But remember to start with the queue, not the technology. Pick a recurring business process where a question regularly waits on a specialist, the underlying data already exists in Snowflake, the answer influences a measurable decision or action, and elapsed time can be measured today. Baseline the current cycle time, implement the agentic workflow, and measure the difference.
Conclusion
CoCo and CoWork are a capable pair, and the working examples are real. CoCo shortens the distance between an idea and a governed, shippable asset. CoWork shortens the distance between a business question and an answer someone can act on. Together, they close the queue that never shows up in a process diagram but costs the business real time every day.
Capability alone doesn't create value, though. The foundation beneath the agents must be ready, the people above them have to remain in command, and the outcome must be measurable in cycle time, not token count.
Keyrus & Snowflake
Deploying CoCo or enabling CoWork is relatively straightforward. Creating an environment where they consistently produce trusted business outcomes is not. That means assessing semantic readiness, access and governance, identifying high-value operating workflows, establishing consumption controls, and defining how humans review and approve agent-generated work.
This is where Keyrus' work with Snowflake stands out: not simply enabling the AI capability, but preparing the data, governance, and operating model around it so that the capability can move a measurable business outcome.
Our proprietary Human Orchestrated Model™ frames that work across three layers:
Intelligence Foundation. The governed data and shared definitions the agent inherits. CoCo is how this gets built. No agent is more trustworthy than the context beneath it.
Human in Command. People set intent, review, and approve; agents execute. Traceable answers and inherited access control are what make that division real rather than rhetorical.
Performance Steering. Measuring whether the agent moved a business number. Cycle time, not token count.
As a Premier Snowflake Partner, we help organizations adopt CoCo and CoWork in a coordinated manner, ensuring that productivity gains do not outpace governance maturity. Learn more about our partnership or contact us to start your Snowflake agentic rollout.
Snowflake CoCo (formerly Cortex Code) is a context-aware coding agent built natively into the Snowflake platform. It works in Snowsight, a desktop app for macOS and Windows, and a CLI, and understands your specific environment, including schemas, RBAC permissions, governance policies, and data relationships, to generate and modify code and semantic assets through natural-language prompts, with human review remaining important before production deployment.
Snowflake CoWork (formerly Snowflake Intelligence) is a conversational agent that lets business users investigate, analyze, and act on governed Snowflake data using natural language. It runs multi-step research, produces reusable artifacts, schedules recurring automations, and can take action in external tools like Google Drive, Salesforce, and Slack through MCP connectors.
Snowflake Cortex AI is a fully managed service that provides low-latency access to industry-leading large language models (LLMs) and search capabilities directly within the Snowflake Data Cloud. It allows users to perform AI tasks like summarization, translation, and data extraction using standard SQL or Python without moving data outside the security perimeter.
CoCo is the builder's agent: it creates semantic models, code, and governed data assets inside Snowflake's operating context. CoWork is the business user's agent: it consumes those assets to answer questions, run investigations, and trigger actions. CoCo builds the capability; CoWork puts it in the hands of the business.
Yes. Both were renamed on June 2, 2026. Cortex Code became CoCo, a cosmetic rename with no change to functionality or architecture. Snowflake Intelligence became CoWork, which reflects a real expansion in capability, from a conversational Q&A interface into a top-level product that runs investigations, builds artifacts, and takes action outside Snowflake.
CoCo is billed on token consumption across every surface, including Snowsight, which was free prior to Summit 26. Cost scales with task complexity, so consumption grows in step with adoption.
Yes. CoWork initializes under the user's DEFAULT_ROLE and applies existing row-access and column-level security policies automatically. Shared artifacts re-run under each viewer's own credentials, so two people can legitimately see different numbers from the same artifact based on their access.
COPILOT_USER and CORTEX_USER are granted to PUBLIC by default, giving broad account-wide access unless revoked. CORTEX_AGENT_USER is not granted by default and must be assigned explicitly. Revoking CORTEX_USER from PUBLIC also removes access to unrelated Cortex AI functions like AI_COMPLETE and AI_CLASSIFY, so per-user credit limits are generally a better control than broad revocation.
Yes. A team can export the existing Power BI model and use CoCo to analyze and reverse-engineer the available semantic-model assets, generating a Snowflake semantic view and flagging objects that cannot be mapped directly. From there, CoCo can create a Cortex Agent against that view for use in CoWork, in the worked example described above, the process took roughly one hour from export to a working conversational agent.
CoCo in Snowsight is generally available to commercial accounts as well as FedRAMP Moderate and High, DoD, and KSA sovereign accounts. CoCo Desktop and the CLI are limited to commercial accounts only. All surfaces require cross-region inference to be enabled, which has data residency implications worth addressing early in any pilot.
The Human Orchestrated Model™ (HOM™) is Keyrus' proprietary structured framework to design, operate, and scale intelligence within organizations. HOM™ is built on three interdependent layers: 1. Intelligence Foundation 2. Human in Command 3. Performance Steering Three layers that turn data, agents, and human governance into measurable performance. Data enables intelligence. Governance directs it. Measurement improves it.
Snowflake provides access to a variety of high-performance models from providers including Google (Gemini), Meta (Llama), Mistral AI, Anthropic (Claude), and OpenAI (GPT-4). This allows users to choose the right model size and performance level for their specific use case.