Only 11% of organizations have agents in production despite 38% piloting them, according to Deloitte's 2026 Tech Trends report. We see this gap between pilots and production deployments happening across industries and companies of all sizes: Teams build a compelling demo and deploy it against real data, only to have it confidently return the wrong revenue number or miss a policy constraint that a junior analyst would have caught.
The problem is a lack of context. Agents need to know which metric definition is authoritative, which data is current, what permissions each user holds and what happened in the last conversation. Without that grounding, even the most capable model operates on guesswork. And even when agents get the answer right, it can be difficult to share those insights with the rest of the business.
At Expedition 2026, Snowflake is announcing capabilities across three dimensions that strengthen how agents and apps run alongside your data, with full governance and context:
- Ground: Connect agents to your governed data, metrics and business logic so they answer questions from authoritative definitions, not guesswork.
- Act: Allow agents to execute real work following each user's own permissions, in the tools teams already use, with clear auditability. Quickly and securely ship Snowflake apps on governed data so people and agents can extract insights and take action.
- Compound: Share proven skills, profiles and memory across the organization so every new agent starts off better than the last.
Ground: Give every agent relevant data and business context from Day 1 #
The a16z research team put it plainly in their March 2026 analysis: "A crucial reason agents didn’t work well was a lack of proper data context." An agent answering the question "What was revenue growth last quarter?" can write competent SQL against a plausible-looking table yet return wrong numbers with full confidence because nobody told it which field carries the authoritative definition of “revenue.”
Snowflake's approach starts from the principle that agents should inherit the context your organization has already built: the schemas, metrics, business logic, governance policies and permissions that define how your data works. Several new and enhanced capabilities make this practical.
Snowflake Cortex Sense (public preview soon) autobuilds context from your existing schemas, metric definitions and business logic, so you don’t spend weeks setting up and defining connections. Cortex Sense works across your own data, shared data and data you purchase through Snowflake Marketplace — so agents understand the entire picture, not just one team's slice of it.
When agents need to reason over large volumes of unstructured information, Analytical Search (public preview) lets them count, compare and find trends across thousands of documents at once, from contracts and support tickets to internal reports. This is not keyword search; Analytical Search applies AI functions across document collections to answer questions such as "How many contracts contain a price escalation clause?" or "Which support tickets mention the same root cause?" And because Cortex Search now enforces row access and masking policies on source tables (private preview), the documents an agent can see follow the same rules as the structured data it queries.
As agents move from finding documents to understanding them, Snowflake’s new embedding model Arctic Embed 3S (generally available soon) improves retrieval quality over Arctic Embed L-v2.0 by 3% in English, 6% in multilingual retrieval, and 23% for structured tables and columns, while keeping the same compact size class.
As frontier models become commoditized, the competitive advantage shifts to how well a model understands your specific business. Organizations are turning to open-weight foundation models that are customized to their specific domain using post-training methods such as supervised fine-tuning and reinforcement learning. Snowflake Cortex Training (public preview) makes it easy to posttrain and deploy open-weight models — Qwen, DeepSeek, Mistral, Gemma and others — on proprietary data using fully managed GPUs. Benchmarks show up to 61% lower cost than competing serverless offerings.<sup>1</sup>
With Cortex Training, teams can train, experiment, evaluate and deploy customized models on the same platform as their governed data. Cortex Training also integrates Cortex AI Gateway (public preview) to capture agent traces and logs, creating a continuous feedback loop for model customization and improvement.
New capabilities make shared and external data agentic as well. The AI Data Package (public preview), with more than 60 curated external data sets across key industries, comes complete with semantic views and a ready-to-use Snowflake CoWork agent as a managed service. Agentic Marketplace Discovery (public preview), powered by Snowflake CoCo, lets users source data, apps and AI through conversation. Agents for Shared Data (public preview) automatically generates a Cortex Agent and semantic view for any data shared, so recipients can use it conversationally. Together, these capabilities enable users to discover, query and use shared and external data entirely through natural language.
Act: Hand agents real work, wherever teams work #
Grounding agents in the right data is the starting point. The next question is whether they can do real work — autonomously, with the right permissions, in the places where teams spend their time. Capgemini's research found that trust in fully autonomous AI agents dropped from 43% to 27% in a single year amid concerns about data privacy, the “AI black box” opacity and limited data readiness.
Snowflake’s integrated, governed approach to AI is designed to prevent that situation, and our latest enhancements help build agent trust through transparency, permissions enforcement and auditability.
Snowflake CoCo: The coding agent that keeps working after you log off
Snowflake CoCo, the coding agent for data teams, can now execute real work autonomously across sessions, schedules and tools.
CoCo Automations (generally available in the Snowsight interface and CLI; public preview in CoCo Desktop) take recurring work off your to-do list by turning it into scheduled Snowflake tasks. A data engineer describes a job in plain language, such as “check pipeline health before standup” or “rebuild the Monday report,” and sets the cadence. Each run produces a transcript, operates within the creator's permissions and runs within preset quotas that cap what an agent can spend. Teams can share automations and the CoCo skills they have built through an account-wide Skills and Plugin Catalog, so a workflow one engineer creates becomes available to the whole team.
Snowsight now has a CoCo mode, a full-screen workspace backed by its own cloud container with a shell, file system, Git and live preview. Start a long migration, a multistep analysis or a pipeline build, close your laptop, and return to finished results. With Cloud Agents (generally available), teams can run CoCo in an isolated, Snowflake managed environment that supports long-running tasks. Every session uses the creating user's Snowflake permissions, and restricted session scope (generally available) lets administrators further limit what each agent can do.
CoCo also now meets teams where they already work. The CoCo Slackbot Integration (public preview) brings CoCo into Slack channels for questions, workflow triggers and result reviews. The CoCo GitHub Integration (public preview) routes proposed code changes through the team's existing pull request process. CoCo in Snowflake Mobile (iOS generally available; Android public preview) lets users check on a cloud session, approve a tool call or send a follow-up from their phone. And with the Cortex Agent SDK (public preview), developers can embed CoCo's coding capabilities inside any application they ship.
For more details, read our blog post about the new CoCo enhancements.
Snowflake CoWork: A personal work agent for every team
Snowflake CoWork is the personal work agent for knowledge workers across an organization. At Expedition, we are adding the ability to tune CoWork to your team’s specific work.
It starts with CoWork Profiles (public preview), which let administrators configure which agents, skills, connectors and instructions each team sees. A finance team's CoWork experience can route questions to a financial planning agent, while the marketing team's CoWork routes to a campaign analytics agent — all without widening access. The same question reaches the right expertise automatically, so users ask questions naturally instead of picking from a menu of agents they do not recognize.
Next, Snowflake Dashboards (public preview) enable anyone on the team to talk to their dashboard through CoWork: summarize it, ask what drove a change or filter to a segment. Analysts build the dashboards by describing them to CoCo and sharing them with the correct roles. The data behind the dashboard stays in Snowflake, governed by existing row access and masking policies.
The dashboards create a shared source of truth that enables teams to have AI-powered conversations that pull from governed data and shared context. Snowflake is welcoming the Mora team to help bolster these efforts by incorporating Mora’s inventive approach to AI for BI into Snowflake’s platform.
For needs beyond dashboards, such as custom exploration, advanced visualizations or multistep workflows, teams can create Streamlit in Snowflake apps or React and Next.js apps on Snowflake App Runtime with built-in conversational AI. CoWork's broader capabilities are also expanding at Expedition. Automations run recurring work on a schedule with the same permissions as a live question. User Memory (generally available) and User Skills carry context from one session to the next so no one re-explains how the team works. MCP connectors, code execution and document generation (PDF, PPTX, DOCX, HTML) turn answers into shareable work products, and HTML Reports (public preview) enable teams to publish polished, governed reports directly from a CoWork conversation.
For more details, read our blog post about the new CoWork enhancements.
Snowflake Bots: Always on, always on it
Snowflake Bots (private preview soon) bring proactive, goal-driven agents to CoCo and CoWork. Give a bot a responsibility, such as keeping critical pipelines healthy or staying on top of customer churn, and it owns the goal from start to finish: it suggests which signals to monitor, plans the work and sees the task through over days or weeks on your governed data, under your existing permissions.
The bot brings you in only when a decision needs human input and reports back when the job is done, so your team can spend less time tracking work and more time on the strategic, creative work only humans can do.
Apps that are reachable by agents and embeddable beyond Snowflake
New enhancements give Snowflake apps, whether internal enterprise apps or customer-facing apps, a broader reach, more flexibility and native support for agentic workflows. Streamlit in Snowflake apps gain several capabilities that expand their reach and conversational abilities:
- User-based sharing (generally available): Teams can share apps with individual users in addition to roles.
- Embedding support (generally available): Streamlit apps are nowembeddable in portals and third-party tools such as Salesforce without requiring a Snowflake login and while preserving governed access.
- Discoverability and viewability: Apps can be surfaced as CoWork Artifacts in CoWork (private preview).
- Conversational capabilities: Users can have side-by-side conversations with CoCo within Snowsight (generally available) and CoWork (private preview) about data in their apps.
- Automatic MCP exposure (public preview): Every app turns into a set of agent-callable tools, invoking app data and actions directly.
For full-stack web apps, App Runtime is an end-to-end solution for going from prompt to deployed web app inside your Snowflake account. Describe your app idea to Snowflake CoCo and it can build and publish it inside your Snowflake account, with SSO, RBAC and audit logging already turned on. Serverless compute (public preview soon) enables apps to scale to zero when idle and wake in seconds when someone opens them, so teams pay for actual usage instead of always-on compute. Workspaces integration (public preview soon) lets teams build and deploy entirely inside Snowsight when local dev tools are restricted. With the Cortex Agent SDK, app.yml v2, deployment targets and horizontal scaling, App Runtime can provide a clear path from CoCo-generated prototype to production operations.
Enhancements to Snowflake Native Apps also enrich what ISV partners and app providers can build. Using App Runtime as a frontend for Snowflake Native Apps (public preview soon) gives providers a full-stack web framework for advanced data exploration and complex workflows. Snowflake Native Apps also now have access to Snowflake Postgres (public preview soon), so providers can create and query Postgres instances directly in account without external access integration workarounds.
All of these capabilities combine to make Snowflake apps a crucial part of the agentic enterprise. Both internal and external-facing apps are now reachable by agents, embeddable in environments beyond Snowflake and built on a foundation of well-governed data.
Compound: Turn individual agent wins into organizational advantage #
The organizational payoff of well-grounded agents comes when the next agent builds on everything the last one learned, and proven workflows, team context and cost-efficient model routing are shared by default rather than rebuilt for every project.
Skills and Plugins Sharing (generally available) turns one expert's workflow into a reusable, governed asset across your Snowflake account. Build a pipeline-review skill or data-quality skill once and share it as a snow:// link or choose to make it discoverable in the searchable Snowsight catalog. Teammates with access can install it in CoCo Desktop, Snowsight CoCo or the CoCo CLI. Admins certify trusted versions, pin what agents use and track adoption. Your teams spend fewer tokens re-solving the same problem and get more consistent outputs, and your agents follow reviewed playbooks instead of improvising.
Agent Profiles bring team-specific defaults — such as instructions, connectors and preferred data sources — into every session. Combined with Memory, which persists personal context across sessions, agents start each conversation with full awareness of how the team works, what was discussed last time and what conventions matter. No one re-explains the tech stack or the naming conventions every Monday morning.
Model Tiers with automatic routing match each task to the right model for the job. Routine summarization and formatting tasks run on cost-efficient models while complex reasoning tasks route to more capable ones. Teams get the quality they need without manually selecting models or overspending on simple tasks.
Finally, Agentic Marketplace Discovery (public preview) brings CoCo to Snowflake Marketplace in Snowsight. Users can describe what they need, such as "company firmographic data for EMEA accounts," and refine it through conversation as CoCo finds and compares matching data, apps and AI products. From the same conversation, users can explore a listing, install it, start a trial or request a quote, when the listing supports it and their role allows. Discovery becomes the first step toward putting new data to work.
Build and use agents that know your business and do real work #
These announcements show Snowflake’s commitment to a common design principle: Agents should run on the same platform as the data they need, inherit governance that already exists and improve with use. When agents are grounded in your data and have the appropriate guardrails to act with real permissions, they can deliver trusted results and responses without adding friction to your AI workflows.
Explore more Snowflake Expedition announcements at snowflake.com/en/blog/expedition-26/, or join the virtual event at snowflake.com/en/expedition/.
Forward-looking statements: This article may contain forward-looking statements, including about future product offerings, and is not a commitment to deliver any product offerings. Actual results and offerings may differ. See our latest 10-Q for more information.
<sup>1</sup> Cost comparison methodology: Pricing for Fireworks and Tinker is based on pricing published on Oct. 22, 2026, on their webpages. We ran LoRA SFT training workloads on Cortex Training with various context lengths, and measured the throughput per H200 GPU. We used the throughput (toks/sec) to calculate price per million tokens for Cortex Training using the formula (price_per_second / throughput) x 1M.
Frequently asked questions #
What agent-related announcements were made at Snowflake Expedition 2026?
Key announcements include CoCo mode in Snowsight, Cloud Agents for autonomous coding workflows, CoWork Profiles for team-specific agent routing, Snowflake Dashboards in CoWork, Cortex Training for customizing open-weight models, App Runtime with serverless compute, Skills and Plugins Sharing for organizational reuse, and Agentic Marketplace Discovery for conversational data sourcing.
How does Snowflake make AI agents more reliable?
Snowflake grounds agents in governed data through Cortex Sense, which auto-builds context from schemas, metrics and business logic. Agents inherit existing permissions, row access policies and masking rules — so they can answer from authoritative definitions rather than guessing. Agent actions run under the user's or owner’s Snowflake permissions and are auditable.
What is Cortex Sense?
Cortex Sense (public preview soon) automatically builds business context from your existing schemas, metric definitions and business logic. It is designed to eliminate the weeks of manual prompt engineering typically required to make an agent understand your data, and works across your own data, shared data and Marketplace data.
What is CoCo Agent Mode?
CoCo Agent Mode gives Snowflake's coding agent a full-screen workspace backed by a cloud container. Data engineers can start long-running migrations, builds or analyses, close their laptop and return to finished results. Sessions run under the user's Snowflake permissions with restricted session scope for additional controls.
What are CoWork Profiles?
CoWork Profiles (public preview) enable administrators to configure a tailored agent experience for each team by specifying which agents, skills, connectors and instructions are available. The same question automatically routes to the right expertise without widening data access.
Does Snowflake have an app development platform?
Yes. You can build and ship enterprise-grade apps on Snowflake, including data exploration apps, full-stack web apps and external-facing data products. Streamlit in Snowflake is used for building internal Python-based data apps such as data dashboards and visualizations, exploratory analytics and lightweight workflows. Snowflake App Runtime allows you to build full-stack apps with custom UIs, including advanced data exploration apps, complex workflow apps and AI apps/agents. The Snowflake Native App framework is for building external-facing, monetizable data products, which can be listed and distributed on Snowflake Marketplace.
How is App Runtime different from Streamlit in Snowflake?
Streamlit in Snowflake is Python-first and designed for data dashboards and visualizations, exploratory analytics and lightweight workflows. App Runtime is for full-stack web apps: advanced data exploration, complex workflows, and AI apps and agents that need custom UI and UX, specialized JavaScript libraries, tailored navigation or API endpoints. Both run inside your Snowflake account and inherit its governance framework.