What's New in AI at Snowflake
AI is no longer a promise. It is creating real opportunity today, and the shift is fundamental: work is becoming a partnership with intelligent agents that operate across your business, continuously and autonomously. This is the agentic enterprise, and these are Snowflake's new and upcoming AI features, grouped by functional area.
How to read this doc
Each tab covers one functional area. Every feature carries a status tag so you can see at a glance what is available now and what is coming. Grouping is for readability only; nothing here is ranked by priority. The final tab, Availability Summary, lists every feature and its status on a single page.
Roadmap items are directional and subject to change. For anything not generally available, the tag reflects its current state.
The agentic enterprise, in four parts
To securely and effectively scale intelligent systems, an agentic enterprise rests on four core components. The AI Data Cloud delivers them as one unified platform, not four products bolted together.
What is in this doc
Seven functional areas, then a summary table. Each area is a tab above.
| Area | What it covers |
|---|---|
| Snowflake CoWork | The business-user AI workspace, formerly Snowflake Intelligence. Personal Agent, Autonomous Agents, Dashboards and Collaboration, Cortex Sense, Agent Studio. |
| Snowflake CoCo | Cortex Code, the AI coding agent for data. Surfaces, Desktop, Cloud Agents, Skills and Plugin Sharing, the Agent SDK. |
| AI Models and ML | Train and fine-tune models on your data. Cortex Training, Cortex AI Function Studio, Online Feature Store, multimodal and document AI functions. |
| Horizon Context | The AI context and semantic layer. Semantic Views, Autopilot, Semantic Studio, Metadata Connectors, Universal Search, the Open Semantic Interchange. |
| AI Security and Governance | Intent-Driven Governance, Sensitive Data Classification, the AI Governance Dashboard, AI Guardrails, AI Security Posture Management, Agent Identity, Data Exfiltration Protection. |
| AI-Powered Data Engineering | CoCo at every layer of connectivity, Snowpipe Automatic Setup, AI-Assisted Troubleshooting, AI-Native Observability, Snowflake AIM, protecting sensitive data at ingest. |
| AI in Sharing and Marketplace | Auto-gen Agents for Data Shares and Listings, Cortex Agent Sharing, AI-powered Marketplace Discovery. |
| Availability Summary | One table of every feature with its status. |
One thread through all of it
Every area here is built on the same governed foundation. Agents reason over your data, act within your systems, and scale with the trust the enterprise requires. The pieces are designed to reinforce each other rather than drift apart.
Snowflake CoWork
Snowflake Intelligence is now Snowflake CoWork. The name evolved and the vision sharpened: one AI workspace where every enterprise context comes together and work gets done. CoWork reached general availability on November 4, 2025, and several capabilities have gone GA or Public Preview since.
How CoWork works under the hood
CoWork is a ready-to-use agentic application powered by Cortex Agents. When a user asks a question in natural language, an orchestrator LLM interprets intent, selects the right tools, and plans the sequence of actions. It may use one tool, chain several together, or decide the question is out of scope. Tools run and return results, then the orchestrator refines and generates the final answer, including summaries, tables, or charts.
- Cortex Analyst: turns natural language into SQL against your semantic views for structured data.
- Cortex Search: searches unstructured data to return relevant document text.
- Custom tools: run user-defined functions or stored procedures to take actions.
Built-in visualization supports most Vega-Lite chart types (area, heatmaps, box plots, dual-axis, layered, faceted small-multiples, error bars, and text annotations, plus bar, line, pie, and scatter). Geographic map charts are not supported. Administrators can use existing identity providers to give teams access only to CoWork, so users interact only with the data experiences built for them.
One workspace. Every context. Any outcome.
CoWork is your chief of staff, turning context into clarity and clarity into action. It brings company data, enterprise context, email, team conversations, apps, presentations, and dashboards into one governed place, then turns them into self-service insights, agentic work, and conversational intelligence.
Personal Agent GA
The Personal Agent is the entry point for every knowledge worker. It knows all of your enterprise data, routes automatically to the right tools and sub-agents, and remembers and learns across sessions. It follows you to mobile and the tools you already use.
Memory, Skills, and Prebuilt Plugins
Three capabilities make the Personal Agent feel custom-built for each role, without anyone starting from zero.
Your Agent, On the Go GA
An iOS app with Face ID unlock, complete conversation history on mobile, and the full CoWork experience on a phone. The app is available on the App Store and supports the same features as the web app: chat, agents, file and image attachments, voice input, citations, and role and warehouse selection. Think of it as the morning briefing that greets a leader before they open the laptop. Generally available July 13, 2026.
Automations Coming Soon
Schedule briefings in plain language, set once conversationally, and they are delivered automatically. Every delivery includes a full AI analysis of what changed, why it matters, and suggested next steps. A follow-up question is one click away, with full context preserved. It is enterprise-governed by default, running with RBAC, cost visibility, and a full audit trail.
Autonomous Agents
Autonomous Agents go beyond answering. They produce finished deliverables, run proactive checks, auto-learn business context, and do deep, analytical research.
Answers dashboards cannot reach Coming Soon
For structured or unstructured, strategic or operational questions, one fully cited answer.
A measurable lift
In CoWork's Hybrid Deep Research Benchmark, this multi-agent approach delivered a 52.2% relative improvement over single-agent systems.
Cortex Sense Private Preview
The managed context runtime that lifts CoWork's accuracy by auto-learning your business from query history, metadata, and BI connections, with zero manual ontology setup. In Snowflake's testing it delivered 3.5x higher accuracy on enterprise business questions compared with Claude Code. Cortex Sense spans CoCo, CoWork, and Cortex Agents, so the full treatment lives in the Horizon Context tab.
The agent does not just answer, it produces Private Preview
A Code Execution Tool runs secure Python inside CoWork with filesystem and network isolation, so your data never leaves the Snowflake perimeter. The output is finished deliverables: PDF reports, PPTX presentations, advanced analysis, and custom documents, all produced under governed, autonomous agent execution.
One Agent. Every Question. Right Tools. GA
Inside each user's existing permissions, a Personal Agent routes to specialized sub-agents: Sales, Finance, Ops, Product, and more. Each can produce PDF reports, PPTX presentations, advanced analysis, and other deliverables through Skills.
Multi-agent routing is real, and refining
An LLM-powered router (the "Personal Work Agent" or Main Agent) automatically dispatches a question to the right specialized agent without manual selection. It is in Private Preview and being refined for latency and routing quality, so for now some teams guide users to pick the right agent explicitly while the router matures.
Dashboards and Collaboration
One finding becomes everyone's starting point. CoWork turns a single answer into shared, live, governed artifacts the whole team builds on.
Artifacts: Live Dashboards GA
Multi-tile Artifacts for Northstar Dashboards Coming Soon
Build on CoCo, publish on CoWork. Analysts create on Workspaces with SQL or natural language through CoCo. Publishing on CoWork is privileged, so you govern what knowledge workers consume. Sharing is role-based, and every dashboard supports high-quality conversational follow-up questions.
Conversation Sharing Coming Soon
Collaborate around real answers with full citations included, no more static screenshots.
Agent Studio: build, manage, and govern every agent in one place
One pane of glass for every agent in your organization. GA
Where to start today
The platform is ready. Start building. Try the CoWork quickstart to deploy your first agent today, read the Snowflake CoWork blog, and use Agent Studio to go from prototype to production in hours.
Snowflake CoCo Cortex Code
Snowflake CoCo is the AI coding agent for data. It moves you from "I have an idea" to "it's deployed" fast, turning data engineering, machine learning, and agent-building tasks into simple conversations. It is deeply aware of enterprise data, compute, governance, and operational semantics across the stack, and it is governed by Snowflake's RBAC wherever you build.
The data-native agent harness
CoCo is built on a live connection to Snowflake, with a custom system prompt focused on data operations rather than general world knowledge. It carries more than 100 domain-specific skills (data governance, machine learning, Iceberg, dynamic tables, cost intelligence, and more), custom tools tuned for data tasks like SQL execution and dbt, and snowflake-hosted LLMs that keep everything inside the security perimeter.
State of the art for data engineering
CoCo outperformed leading coding agents on ADE Bench and pulled even further ahead on real Snowflake dbt project tasks. ADE Bench is a framework created by dbt for evaluating AI agents on real-world analytics and data engineering tasks.
Use CoCo where you work today GA
Four core surfaces are available today.
CoCo Desktop GA
The newest surface: a native, VS Code-based application for full agentic development on data. CoCo Desktop reached general availability on July 21, 2026 (v1.20.0 shipped July 22).
Cloud Agents power CoCo in Snowsight GA
All the power of the CoCo CLI with the same robust agent loop, skills, tool execution, and runtime, securely inside Snowflake.
Cloud Agents enable bash and Python script execution, web search, file read and write, and dbt execution.
Interact with CoCo on Slack or on the go
CoCo Slack Bot Coming Soon
Ask questions, kick off workflows, review outputs, and unblock tasks, all from Slack. Get a quick notification, allow CoCo to continue an action, and stay unblocked without sitting at the keyboard.
CoCo Mobile App Coming Soon
The same API that powers Cloud Agents enabled a CoCo mobile app, so you can keep work moving from a phone.
Skills and Plugin Sharing Private Preview
Individual productivity compounds into organizational leverage. A top customer ask, now in private preview.
CoCo Agent SDK NEW GA
CoCo is now an agent platform with an SDK. The same engine that powers CoCo for thousands of customers is now an installable library you can embed in your own applications and workflows.
The SDK packages CoCo's tools and agent loop with built-in tool execution (Read, Write, Edit, Bash, Glob, Grep, SQL), sessions, streaming, hooks, structured output, and MCP, so your agent can start working immediately without you implementing that infrastructure yourself.
Build on CoCo programmatically GA
Model Context Protocol (MCP) Server
Interact with CoCo programmatically through an MCP server, so other agents and tools can call CoCo for data tasks.
Agent Client Protocol (ACP) Support
ACP support lets you integrate CoCo into broader agent client workflows.
AI Models and ML
AI and ML run where your data lives, so teams train, fine-tune, and operationalize models without moving data out of Snowflake. Snowflake ML covers the full lifecycle in one platform: prepare data, build features, train on CPU or GPU, track experiments, register and serve models, then monitor them in production. The newest shift is agentic ML, where CoCo drives that lifecycle from natural language prompts.
Agentic ML with CoCo GA
The headline shift in Snowflake ML: CoCo autonomously plans, executes, and iterates on machine learning workflows on your behalf. From a natural language prompt, it can explore data, engineer features, train and evaluate models, debug issues, and prepare models for deployment. You can hand it a single task or a broader goal that needs multi-step reasoning, with native context awareness across your data, models, notebooks, and features.
Why this matters for a data science team
The bottleneck in enterprise ML has rarely been the algorithm. It is the infrastructure work around it: moving data, standing up environments, wiring pipelines, keeping training and serving consistent. Agentic ML collapses that work into prompts against governed data that never leaves Snowflake.
Cortex Training Private Preview
Fine-tune open-weight models such as Qwen and Mistral on your enterprise data inside Snowflake, without moving that data out. Use it for domain adaptation and reinforcement-learning-based improvement, so a general model learns your business rather than a generic one.
Training infrastructure
Cortex AI Function Studio Public Preview
Create, evaluate, and optimize Cortex AI Functions in one place. Public Preview as of May 20, 2026. It turns the loop of writing an AI function, testing it against your data, and tuning the prompt or model into a guided workflow rather than a manual, trial-and-error exercise.
The AI Function library GA
Run unstructured analytics on text and images with models from OpenAI, Anthropic, Meta, Mistral, and DeepSeek, all deployed inside the Snowflake service perimeter. The core set is generally available and callable from SQL or Python.
Two practical notes worth raising
AI Functions are optimized for throughput, so they are the right tool for batch work over large tables. For interactive, latency-sensitive paths use the REST APIs (Complete, Embed, Agents) instead. On access: calling them requires the USE AI FUNCTIONS account privilege plus the CORTEX_USER or AI_FUNCTIONS_USER database role, and you can now grant per-function privileges such as USE AI FUNCTION AI_COMPLETE for finer control.
Multimodal and document AI functions
Cortex AI Functions keep expanding across modalities and document types. A few that landed in 2026:
| Capability | What it does | Status |
|---|---|---|
| Multimodal video and audio analysis | Marketing and brand insights from video and audio, including Gemini 3.5 Flash support in Cortex AI Functions. | Public Preview |
| AI_CLASSIFY document classification | Classify documents by type, not just rows of text. | Public Preview |
| AI_COMPLETE document intelligence | Document-level intelligence and extraction with structured outputs. | Private Preview |
| Fine-tuning arctic-extract models | Fine-tune arctic-extract for domain-specific document extraction so it handles your forms and contracts better. | Private Preview |
| AI_EXTRACT extraction scores | Confidence scores on extraction results so you know what to trust. | GA |
| AI_TRANSCRIBE AAC audio | Transcription now supports AAC audio files in addition to existing formats. | Public Preview |
Online Feature Store Public Preview
A production feature store for real-time ML, Public Preview as of July 10, 2026. Define features once, serve them online at low latency, and keep training and serving consistent so the features a model trained on are the features it scores with. It is backed by a managed Snowflake Postgres serving layer and built for fraud detection, recommendations, and personalization.
Three ways to define a feature
| Feature view type | When you use it | Status |
|---|---|---|
| Batch | Features computed on a schedule and synced to the online store. Serves the latest row per entity key. | Public Preview |
| Stream | Ingest events in real time through a REST endpoint and serve continuously updated features, including rolling time-windowed aggregates such as spend over 48 hours. | Public Preview |
| Real-time | Compute at query time from per-request inputs plus upstream features, for values that cannot be precomputed such as the current transaction amount or live location. | Public Preview |
Feature groups bundle multiple feature views into one versioned set, so a model fetches everything in a single round-trip and the same group can generate the training set. Point-in-time correct training joins keep future data from leaking into a model.
Model Registry and serving GA
Models are first-class Snowflake schema objects, so they inherit the governance you already run: RBAC, tags, roles, and sharing. Log models trained anywhere, then run inference at scale from Python, SQL, or a REST endpoint, in a warehouse or on Snowpark Container Services for GPU serving. Built-in support covers scikit-learn, XGBoost, LightGBM, PyTorch, TensorFlow, Keras, Prophet, CatBoost, HuggingFace pipelines, and MLflow, plus your own custom code.
Close the loop: features feed inference automatically
Deploy a registered model as a real-time inference service and map a feature view to a model method. At inference time the service fetches any missing feature columns from the online store itself, so your application can send just the entity IDs and let Snowflake retrieve the rest.
ML Observability, Explainability, and Lineage GA
Models degrade quietly. Input drift, stale training assumptions, and pipeline changes do not throw errors, they just make predictions worse. ML Observability tracks production model quality across performance, drift, and volume for regression, binary, and multi-class classification, and metrics are queryable through SQL table functions so you can build your own dashboards alongside the Snowsight views.
Why it matters for the conversation
CoWork and CoCo are how users and builders work with AI. This tab is how teams bring their own models and functions to that experience: fine-tune a model on your data, package an AI function, serve features at low latency, and apply it to documents, audio, and video, all without the data leaving Snowflake.
Horizon Context
Context exists in your organization, but it is fragmented and inactive. Your schemas are in Snowflake, your query patterns are in query history, your transformation logic is in SQL and dbt, and your BI definitions live in Tableau and Power BI. None of that flows automatically to your AI agents today. Horizon Context changes that through three phases: Collect, Enrich, and Activate.
Semantic Views GA
One definition, now more powerful, and far easier to build. Semantic Views are how Horizon Context reaches the agent. Your tables, joins, metrics, and filters are described once in a governed model, and that definition becomes the single source of truth that Cortex Analyst, Power BI, and your custom apps all resolve against. One version of the truth, everywhere.
Semantic Views are now significantly enhanced: complex calculations, composable definitions you reuse across models, and automatic query rewrite so Horizon picks the most efficient path without the agent having to know.
Semantic View Autopilot GA
The bigger announcement is how much easier it is to build Semantic Views. Give Autopilot your existing SQL queries, your Tableau workbooks, your Power BI reports, the logic your analysts have already built and validated, and it generates a complete semantic view automatically. Your tribal knowledge becomes a governed semantic layer, without anyone starting from scratch.
Autopilot proposes. Humans approve. AI executes with confidence.
The generated view is a proposal a person reviews and approves before it goes live, so governed definitions stay human-controlled.
Snowflake Semantic Studio Private Preview
A purpose-built IDE for building, testing, and deploying semantic views, built into Snowsight.
Built for complexity
Enterprise data models require more than simple star schemas. Semantic Views now support the patterns real businesses have.
| Capability | What it handles | Status |
|---|---|---|
| Semi-additive metrics | Metrics that aggregate across some dimensions but not others, like account balances that sum over geography but take the last value over time. | GA |
| Ratio and derived metrics | Calculates numerator and denominator at different grains before dividing, like revenue per employee. | GA |
| Level-of-detail | Automatically calculates at the correct grain regardless of query context. | Private Preview |
| Range relationships | SCD Type-2 support via BETWEEN and EXCLUSIVE joins on start and end dates. | GA |
| Composable views | A new IMPORTS section to reuse dimensions across different Semantic Views. | Private Preview |
| Many-to-many | Complex associations using intermediate bridge tables. | Private Preview |
Integrated with Horizon Governance
Model-level lineage
Model-level lineage for Semantic Views makes AI insights more trustworthy by showing exactly how business metrics are derived.
Tagging in Semantic Views
Tagging support lets teams consistently classify and govern metrics, models, dimensions, and facts across the analytics stack.
Metadata Connectors Private Preview
Select Star technology, acquired by Snowflake, is now native to Horizon Catalog, shipping as out-of-the-box metadata connectors. This is more than ingesting table schemas.
Five connectors are available now in Private Preview: PostgreSQL, SQL Server, dbt, Tableau, and Power BI. Configure one, point it at your system, and Horizon collects the technical metadata: schemas, column names, and asset relationships. It also collects behavioral signals: query logs that show which tables your analysts actually run, dashboard definitions that encode business logic, and BI-to-table lineage that tells you which dashboard depends on which table. Five more connectors are planned for Q3.
Universal Search GA
Universal Search is now 30% faster, covers more asset types (Tableau dashboards, dbt models, external database tables, Iceberg tables in Glue), and has smarter relevance that understands asset relationships, not just keywords. One search box, every connected system.
Agentic catalog discovery
The bigger announcement: ask Cortex Code in plain English, "Show me all tables tagged as EMEA customer data with a churn signal." CoCo searches Horizon Catalog, finds matching assets across every connected platform, and brings them back with schema, ownership, and full context. No SQL, no manual browsing.
Open Semantic Interchange (OSI) GA
To make governed semantics portable across the entire industry, Snowflake co-founded the Open Semantic Interchange. Since launch, OSI has grown 400%, from 16 founding partners to over 80 today, including 26 new partners through OSI for Financial Services. The V1 spec is finalized and live on GitHub under Apache 2.0.
Any agent, BI tool, or analytics engine that supports OSI can read your governed definitions from Horizon Context without translation.
AI Copilot for Horizon Catalog Coming Soon
A conversational interface to the metadata and usage data in your Snowflake account, right in the UI. It lets data governors, stewards, and security admins self-serve without SQL knowledge or familiarity with the account usage views, so governance and security insights arrive faster and depend less on specialized data teams.
Cortex Sense Private Preview
The managed context runtime that continuously builds and maintains a Snowflake-managed knowledge layer from metadata, documentation, query history, lineage, and connected systems. It improves the accuracy of CoCo, CoWork, and Cortex Agents with minimal manual setup, so production-grade agents ship in days, not weeks. In Snowflake's testing it delivered 3.5x higher accuracy on enterprise business questions compared with Claude Code.
The accuracy jump, in numbers
Without a semantic model, CoCo or CoWork answers are around 47% accurate. With Cortex Sense, that climbs to 86%, the level where agents deliver real business value. The cost is small: roughly $0.49 to $0.59 per prompt, for a 3.6x improvement in accuracy. Cortex Sense is fully managed by Snowflake, so there is no infrastructure to run and no semantic layer to maintain by hand.
How it fits with the rest of this tab
Three related but distinct things. Horizon Context is the governed context layer that connects and organizes enterprise context. Semantic Views remain the gold standard for curated, governed domains, the metrics you hand-model and trust most. Cortex Sense is the managed context runtime that extends AI-ready context across the long tail of enterprise data, the questions you did not model in advance. Cortex Sense treats your governed Semantic Views as the authoritative signal and defers to them, so the pieces reinforce each other rather than drift apart.
AI Security and Governance
As agents reach your data, security has to be built in, not bolted on. Snowflake's approach layers agent security on top of the platform and data security it has had for years: authentication, RBAC, network policies, sensitive data protection, encryption, and data integrity, now extended to data exfiltration and ransomware protection, and then to purpose-built controls for the agentic era.
Intent-Driven Governance Coming Soon
Describe the intent. Snowflake implements it. A Cortex Code skill that turns plain-English governance intent into fully deployed Snowflake controls: masking policies, row access, classification, tagging, drift detection, and alerts. Your CISO approves the intent, and Snowflake enforces it.
A sentence, not an engineer
Your CISO says, "We need HIPAA-compliant data handling." That intent, one sentence, goes into Cortex Code. Cortex Code generates a plain-English specification the CISO reads and approves, no Snowflake knowledge required. Before deployment, a plan report shows exactly what will be configured. After, a full audit package traces every configured object back to that spec. Tag-based enforcement covers your entire estate, including tables that do not exist yet. And CoCo monitors continuously for drift.
Built on three governance skills already in production: Data Governance, Data Quality, and Lineage. Intent-Driven Governance is the CISO-level orchestration layer on the same foundation.
AI-powered Sensitive Data Protection
Find it: Sensitive Data Classification Private Preview
150+ LLM-integrated classifiers covering PII and PCI categories detect sensitive data regardless of column names, continuously, as data evolves, and auto-tag detected columns across your entire estate. No infrastructure setup, no manual column-by-column review.
Protect it: Tag-based policies Public Preview
Tag-based Row Access and Projection Policies attach protection to the tag itself, not to individual tables. One policy covers thousands of tables, including ones that do not exist yet. The same tag that governs a Power BI dashboard now governs AI agent outputs, so no special AI policy is required.
Why it matters
Supports compliance requirements (GDPR, CCPA, HIPAA, PCI-DSS) with AI-powered classification that works at scale, automatically. The protection travels with the data: SQL, BI, and AI outputs, everywhere.
AI Governance Dashboard GA
See every agent, skill, and MCP server, and govern them by asking. How many agents are running in your account right now, not deployed but actually running and touching data? Most teams do not know. The AI Governance Dashboard gives you that inventory: every agent, skill, and MCP server, which ones are touching sensitive objects, and the full audit record for every interaction.
Govern your AI estate by asking
"Why is my churn agent returning stale answers?" "Which agents are touching sensitive data?" "Are any agents reading PII without masking policies?" The dashboard identifies governance gaps with immediate, actionable fixes.
Horizon AI Guardrails GA
Prompt injection is one of the top AI security attack vectors. Snowflake provides built-in protection for known prompt injection patterns for all customers, no configuration required. On top of that, Horizon AI Guardrails provide zero-day-style prompt injection protection that enterprises can simply enable at the account level. They cover Cortex Code, Cortex Agents, and Snowflake CoWork, a unified governance layer across all AI surfaces.
AI Security Posture Management (AI SPM) GA
A significant AI security capability, now generally available. Trust Center monitors not just your data but your AI workloads and your agents. A key benefit of the built-in AI Security scanners is that they run within your Snowflake account, so no data leaves your trusted boundary. The dashboard proactively scans for AI-specific risks, assesses compliance posture against emerging regulations, and deploys programmatic remediations to lock down misconfigurations.
Monitor Cortex AI and first-, second-, and third-party agents. Enable via Trust Center, and use it via UI, the CoCo CLI, programmatically, or org-wide.
Agent Identity GA
One of the most fundamental building blocks, now generally available. Today, when an AI agent takes an action in Snowflake, you cannot easily distinguish what the human did from what the agent did. That is a huge auditability gap. With Agent Identity, every agent gets a unique, recognizable principal within the user session. Every action it takes is stamped with that identity: traceable, auditable, and attributable. Dedicated agent identity tracking is also available in Account Usage views for auditing.
Scope what an agent can do
Agent identity also lets you scope agent access to sensitive data, enforcing restrictions like "this agent cannot access masked PII." Use system functions such as SYS_CONTEXT('snowflake$current', 'is_agent_activated') in data governance policies to restrict data access to agents, and use the agent's unique ID in ACCOUNT USAGE views (Query_History and Access_History) to identify actions the agent performed on behalf of the user.
Extends to third-party agents
The same identity framework extends to third-party agents through integrations with security partners like 1Password, Aembit, Cyera, Linx Security, Okta, SailPoint, and Saviynt, so the governance policies you apply to Snowflake-native agents can cover external AI tools too.
Restricted Session Scope GA soon
Limits what an agent session can do to only what the task requires, so a read-only analysis stays read-only even if the user's role normally allows much more.
Advanced Data Exfiltration Protection Public Preview
It is not a question of whether agents will access your sensitive data, but when. Agentic AI significantly increases the risk of data exfiltration. The Data Exfiltration Prevention (DXP) package is in preview via the Trust Center, paired with Data Movement Policies (now GA) that enforce strict data movement controls.
Data Movement Policies GA
Prevent sensitive data from moving outside approved boundaries. Policies are activated on sensitive data tags with query behavior and results, configured via SQL and DDL.
Trust Center Data Exfiltration Detections
Detect and help intercept unauthorized data flows before they exit the ecosystem, including sensitive data fetches triggered by AI agents, unauthorized routing to stages, and mass downloads via UI.
Context-Aware Access Policies Private Preview
A zero-trust control that lets security admins author a single policy evaluating identity, network, and client context jointly in one expression, so access decisions consider the full context of the request, not any one signal in isolation.
Ransomware Protection via Multi-Party Approval (MPA) GA
Removes single points of failure from your most sensitive architecture by requiring two or more authorizations before any destructive system change can proceed. Even if top-tier administrative credentials are hijacked, ransomware actors cannot unilaterally wipe data or alter configurations.
Client-side CoCo CLI VM Sandbox Private Preview
Run AI-assisted development workflows while minimizing exposure of credentials, local storage, or networks to client-side AI workloads. VM sandboxing isolates each CoCo session in a separate Linux kernel, isolated from the host OS, so developers get the full CoCo experience while security teams get the isolation they require. Admin-enforced or self-managed, currently on macOS.
Cortex AI Gateway Private Preview
The connective layer for all trusted agent activity, built on the Natoma acquisition (a centralized MCP gateway that enforces identity, policy, and audit at the tool-call level). Enterprises are deploying agents everywhere, Claude Code, LangChain apps, internal bots, and more, and today there is no central control plane for what those agents can do. Every new agent type requires manual configuration, there is no audit trail, and AI costs are invisible. Cortex AI Gateway is a single endpoint that sits in front of all of it. Agents route through it, and Snowflake enforces who can call what tool, logs every action, and controls spend.
Why it is more than the Natoma gateway it is built on
Governance is tied directly to Snowflake's existing identity and RBAC system, so there is no new security infrastructure to stand up. The model catalog is model-agnostic by design, spanning GPT, Claude, Gemini, Grok, Mistral, and more. Cost controls and intelligent model routing are built in, so simpler tasks route to cheaper models automatically. And every agent trace and tool call lands in customer-owned tables for audit, compliance, and forensics.
It governs how AI agents, both first-party (CoCo and CoWork) and third-party (Amazon Bedrock, Azure AI Foundry, ChatGPT, Claude Code, Cursor, and custom LangChain or LlamaIndex apps), access models, data, MCP servers, and enterprise tools. You can adopt it modularly: stand up the MCP Gateway for tool governance, or the Model Gateway for routing, independently. Agent Identity, which it builds on, is GA today.
Control
Grant, restrict, and audit model and tool access from a single endpoint, instead of manually configuring each new agent type.
Visibility
Agent actions are captured in real time, so security and compliance teams get the evidence they need without instrumenting each agent individually.
Cost and performance
Route requests automatically to the right model and enforce spending limits before costs run away.
AI-Powered Data Engineering
The reason most production AI underperforms is the data: whether it is fresh, governed, and complete with the business context AI needs to operate reliably. CoCo is woven into every layer of connectivity, so your data engineering pipelines are not just faster to build and easier to run, they are AI-ready from the moment they go live. Transformation, orchestration, observability, and migration all run natively, with no separate infrastructure to operate.
CoCo at every layer of connectivity GA
CoCo is woven into every layer of connectivity. Purpose-built CoCo skills understand these specific products, their patterns, and their failure modes. On ADE Bench, CoCo achieved a 72.1% pass rate, outperforming both Claude Code and OpenAI Codex (each at 65.1%), and versus Claude Code on the same model it uses 51% fewer tokens and takes roughly an hour less. Two design choices drive that efficiency: targeted rather than exhaustive exploration, and a SQL-native approach that leans on native execution instead of heavy bash-based workflows.
Snowpipe Automatic Setup with AI Private Preview
Snowflake automates the cloud setup that used to take hours of manual work and troubleshooting. Give your cloud storage path to CoCo and ask to set up a pipe in natural language. Snowflake handles the cloud plumbing end to end: the IAM role creation, the SQS queue configuration, the notification setup, and the pipe creation. Go from raw files in object storage to a live, continuously-ingesting pipe in minutes.
AI-Assisted Troubleshooting in Snowsight Private Preview
When a pipeline fails, CoCo diagnoses the issue and recommends a fix right inside Snowsight, with context from the Openflow observability dashboard. Built-in observability shows what moved, where it went, and when, so pipeline failures are never silent.
dbt Projects on Snowflake GA
The full dbt lifecycle runs inside Snowflake: develop, deploy, orchestrate, and observe your transformations in the same place the data already lives. Snowflake provides managed dbt Core and dbt Fusion runtimes, so there is no Python environment to maintain, no Airflow cluster to scale, and no dbt CLI version drift across developer machines. Executions use a virtual warehouse at standard compute rates, with no licensing or per-user fees.
Where CoCo speeds this up
CoCo is integrated with Horizon Catalog, so it can generate sources.yml and models.yml documentation from catalog metadata that already exists (table comments, column comments, tags), scaffold dbt data quality tests into your schema.yml, and inspect the files of a deployed dbt project object to debug a production run.
Two things worth flagging to a data engineering team
Deploy to production through CI/CD, not from a workspace. The Snowflake CLI runs in GitHub Actions, GitLab, or Azure DevOps with OIDC ephemeral tokens, so there are no long-lived credentials stored as repo secrets, and every pull request is validated with dbt build before merge. Second, permissions are deliberately split: the MONITOR privilege on a dbt project object exposes lineage, run history, and artifacts, but it does not grant SELECT on the tables the pipeline creates. Building a pipeline and reading its output are separate grants.
Practical limits to plan around: one concurrent execution per dbt project object, and up to 20,000 files per object. Semantic views can be codified in the pipeline using the Snowflake Semantic View dbt package, so they are version-controlled and promoted through the same review process as your models.
AI-Native Observability Public Preview
You do not need deep telemetry expertise to self-serve monitoring. Next-Generation Event Tables eliminate setup, access, and scale barriers. They are zero-config, RBAC-governed, and performance-optimized.
AI Observability and Cortex Agent evaluations Public Preview
Pipelines are not the only thing that needs observability once agents are in production. Cortex Agent evaluations let you test, baseline, and improve an agent's behavior so you know when it is actually ready for users. Rather than grading only the final answer, evaluation follows Snowflake's Goal-Plan-Action framework and scores the agent at each stage of its reasoning, so you can pinpoint where it went wrong instead of just seeing that it did.
| Metric | What it measures | Status |
|---|---|---|
| Tool selection accuracy | Goal to plan: did the orchestration layer invoke the tools you expected for the user's goal. Penalizes too few calls, too many, and the wrong ones. | Public Preview |
| Tool execution accuracy | Plan to action: did each tool receive appropriate input and return output that meets your requirements. | Public Preview |
| Answer correctness | Closes the loop back to the goal: how closely the final response matches expected ground truth. | Public Preview |
| Logical consistency | Spans the whole loop: consistency across instructions, planning, and tool calls. Reference-free, so it needs no ground truth dataset at all. | Public Preview |
CoCo runs the evaluation loop, not just the pipeline
CoCo can generate a synthetic evaluation dataset from your agent configuration, import real queries from production monitoring, run the evaluation, investigate failures, and then suggest and test changes that improve the agent's scores. You can also define custom LLM-judge metrics for domain criteria the streamed answer does not expose, such as which tables the agent actually touched.
Compare up to three runs side by side to see how a configuration change moved each metric, and drill into any single record for the full trace: planning, each tool invocation, and response generation. One caveat to set expectations on: evaluations do not currently support MCP server tools, the code execution tool, or skills.
Snowflake AIM: AI-driven modernization Public Preview
Snowflake AIM is the AI-driven modernization and virtualization platform for migrations to Snowflake. Instead of manually untangling complex environments, AIM analyzes code, workflows, and dependencies to create a clear migration plan and identify operational risk early. At the center is the Snowflake CoCo Migration Agent, providing AI-powered code understanding and translation across the process.
Virtualization
Run Teradata workloads directly on Snowflake with minimal SQL rewrites and zero downtime. Escape renewal pressure, keep applications running unchanged, protect mission-critical operations, and modernize on your terms. Powered by Snowflake CoCo.
Modernization
Automate migration of warehouses, ETL workflows, and Spark workloads into Snowflake-native architectures. AIM handles assessment and dependency analysis, automated code conversion, testing and validation, and ETL modernization, with shared migration state across the whole project.
Supported sources include Oracle, SQL Server, Teradata, AWS Redshift, Google BigQuery, Greenplum, Sybase, Synapse, Netezza, PostgreSQL, and Databricks SQL.
Protect Sensitive Data Before It Ever Lands Private Preview
Tokenize sensitive data during COPY and Snowpipe loads so raw values never reach your destination tables. Bring your preferred third-party tokenization service via external functions and apply policies directly to columns. This minimizes post-load ETL for data protection and gives you one consistent protection pattern across both bulk ingestion and Snowpipe.
AI in Sharing and Marketplace
Data products are becoming agentic. Snowflake now makes it effortless to turn your data shares and listings into conversational agents, so consumers ask questions in natural language, explore follow-ups, and discover new use cases right on top of your data.
Auto-gen Agents for Data Shares and Listings Public Preview
Instantly create a Semantic View and Cortex Agent from any data listing's schema, metadata, and query history. Edit and test before publishing, with full governance and first-party enrichment built in. Enterprise security and governance are included, and consumers can seamlessly combine third-party data with their own first-party data.
Cortex Agent Sharing Public Preview
Distribute these agents across internal teams, partners, or the broader Snowflake ecosystem via Marketplace. Consumers add an agent to Snowflake CoWork in one click. Providers you already trust, like Crunchbase and Dun & Bradstreet, have added agents on top of their existing data listings, giving customers a more conversational way to explore their data.
AI-powered Marketplace Discovery Research Preview
A conversational way to find the right data, apps, and AI, powered by Cortex Code.
Why it is unique
Every major cloud platform is moving toward conversational Marketplace discovery, and Snowflake is doing the same. The unique advantage is that Snowflake is the platform where your data lives and where your analysis happens, so discovery can leverage your context to find the right third-party data, AI, and applications for your existing workflows.
Availability Summary
Every AI feature in this doc, with its status. The new and upcoming answer on a single page. Status was verified against Snowflake release notes as of early August 2026, so several capabilities that were previously "coming soon" have since gone GA or Public Preview.
| Feature | Area | Availability |
|---|---|---|
| Snowflake CoWork (the workspace) | CoWork | GA |
| Personal Agent | CoWork | GA |
| User Memory | CoWork | Coming Soon |
| User Skills | CoWork | Coming Soon |
| Prebuilt Plugins (Finance and Sales) | CoWork | Coming Soon |
| CoWork Mobile (iOS) | CoWork | GA |
| Automations (scheduled briefings) | CoWork | Coming Soon |
| Deep Research | CoWork | GA |
| Analytical Search | CoWork | Public Preview |
| Cortex Sense | Horizon Context | Private Preview |
| Code Execution Tool (Python in CoWork) | CoWork | Private Preview |
| Artifacts: Live Dashboards | CoWork | GA |
| Multi-tile Northstar Dashboards | CoWork | Coming Soon |
| Conversation Sharing | CoWork | Coming Soon |
| Agent Studio (build, manage, govern) | CoWork | GA |
| MCP sub-agents (Sales, Finance, Ops, Product) | CoWork | GA |
| CoCo CLI | CoCo | GA |
| CoCo in Snowsight | CoCo | GA |
| CoCo for VS Code | CoCo | GA |
| CoCo Plugin for Claude Code | CoCo | GA |
| CoCo Desktop | CoCo | GA |
| Cloud Agents (powering CoCo in Snowsight) | CoCo | GA |
| CoCo Slack Bot | CoCo | Coming Soon |
| CoCo Mobile App | CoCo | Coming Soon |
| Skills and Plugin Sharing | CoCo | Private Preview |
| CoCo Agent SDK | CoCo | GA |
| CoCo MCP Server | CoCo | GA |
| CoCo ACP Support | CoCo | GA |
| Cortex Training (fine-tune open-weight models) | AI Models and ML | Private Preview |
| Cortex AI Function Studio | AI Models and ML | Public Preview |
| Online Feature Store | AI Models and ML | Public Preview |
| Multimodal video and audio analysis | AI Models and ML | Public Preview |
| AI_CLASSIFY document classification | AI Models and ML | Public Preview |
| AI_COMPLETE document intelligence | AI Models and ML | Private Preview |
| Fine-tuning arctic-extract models | AI Models and ML | Private Preview |
| AI_EXTRACT extraction scores | AI Models and ML | GA |
| Semantic Views | Horizon Context | GA |
| Semantic View Autopilot | Horizon Context | GA |
| Snowflake Semantic Studio | Horizon Context | Private Preview |
| Level-of-detail metrics | Horizon Context | Private Preview |
| Composable Views | Horizon Context | Private Preview |
| Many-to-many relationships | Horizon Context | Private Preview |
| Model-level lineage for Semantic Views | Horizon Context | GA |
| Tagging in Semantic Views | Horizon Context | GA |
| Metadata Connectors (PostgreSQL, SQL Server, dbt, Tableau, Power BI) | Horizon Context | Private Preview |
| Universal Search | Horizon Context | GA |
| Agentic catalog discovery (via CoCo) | Horizon Context | GA |
| Open Semantic Interchange (OSI) | Horizon Context | GA |
| AI Copilot for Horizon Catalog | Horizon Context | Coming Soon |
| Intent-Driven Governance | AI Security | Coming Soon |
| AI-powered Sensitive Data Classification | AI Security | Private Preview |
| Tag-based Row Access and Projection Policies | AI Security | Public Preview |
| AI Governance Dashboard | AI Security | GA |
| Agent Quality Incident Response (via CoCo) | AI Security | GA |
| Horizon AI Guardrails | AI Security | GA |
| AI Security Posture Management (AI SPM) | AI Security | GA |
| Agent Identity | AI Security | GA |
| Restricted Session Scope | AI Security | GA soon |
| Third-Party Agent Identity integrations (1Password, Okta, SailPoint, etc.) | AI Security | Private Preview |
| Context-Aware Access Policies | AI Security | Private Preview |
| Advanced Data Exfiltration Prevention (DXP) | AI Security | Public Preview |
| Data Movement Policies | AI Security | GA |
| Ransomware Protection via Multi-Party Approval (MPA) | AI Security | GA |
| Client-side CoCo CLI VM Sandbox | AI Security | Private Preview |
| Cortex AI Gateway | AI Security | Private Preview |
| CoCo at every layer of connectivity | Data Engineering | GA |
| Snowpipe Automatic Setup with AI | Data Engineering | Private Preview |
| AI-Assisted Troubleshooting in Snowsight | Data Engineering | Private Preview |
| AI-Native Observability (Next-Gen Event Tables, AI-Assisted Alerts) | Data Engineering | Public Preview |
| Snowflake AIM (AI Migration Agent) | Data Engineering | Public Preview |
| Protect Sensitive Data at Ingest (tokenize during COPY and Snowpipe) | Data Engineering | Private Preview |
| Auto-gen Agents for Data Shares and Listings | Sharing | Public Preview |
| Cortex Agent Sharing | Sharing | Public Preview |
| AI-powered Marketplace Discovery | Sharing | Research Preview |
A note on roadmap items
Anything tagged Coming Soon, Private Preview, Public Preview, or Research Preview is directional. Timing may change, and features under development may ship in a different form. For anything generally available, official Snowflake documentation is the source of truth.