Snowflake CoWork Deep Research now cites web sources next to your data

1 hour ago

7 min read

Snowflake CoWork Deep Research web search cover with title over a teal city skyline - Data Studios original cover

Snowflake has extended Deep Research in Snowflake CoWork so that an investigation can also research the public web and cite what it finds alongside the findings drawn from your own data. The change appears in Snowflake's release notes dated October 30, 2026, and it applies only when the agent behind the conversation has the web search tool enabled. When web search is off, Deep Research behaves as before and works only with the data the agent can reach inside the account.

The release note sets out three rules that define the feature. Web findings are cited in the final report next to the findings from your data, the web research runs separately and has no access to your data, and investigations without web search use only your data. For a data team, the practical consequence is that a CoWork report can now combine an internal metric, such as a drop in regional sales, with external context gathered on the web, while the citations keep the two kinds of evidence distinguishable.

Snowflake CoWork is the conversational application formerly called Snowflake Intelligence, and the older name still appears in some product objects and SQL keywords. The release note does not state a preview or general availability label for the web research capability, and the CoWork overview page does not state one for the product either, so the status should be confirmed in the account before the feature is rolled out to business users.

··········

HOW SNOWFLAKE COWORK DEEP RESEARCH WORKS.

Deep Research is an investigation mode in CoWork intended for complex, open-ended questions that need several steps. According to Snowflake's documentation, it breaks the question into sub-investigations, runs them in parallel across structured and unstructured data, and synthesizes the results into a structured report in which every claim is traced to its source data and queries. A user starts it by selecting the plus button in the message bar and choosing Deep Research. Investigations can take up to 10 minutes, and the finished report stays in the thread context, so follow-up questions run as normal chat turns.

The documentation recommends Deep Research for causal "why" questions that span multiple tables or business domains and for reports that need citations before they go to stakeholders. For quick lookups and single metrics, standard chat is the suggested mode. Snowflake also separates Deep Research from extended thinking: extended thinking applies deeper reasoning to a single answer, while Deep Research performs broader autonomous exploration, and the two can be combined in the same conversation.

Underneath, CoWork sends each question to the Cortex Agent API. An orchestration model interprets the request, selects tools and plans the steps. The documented tool types are Cortex Analyst, which generates SQL against semantic views for structured data, Cortex Search, which retrieves relevant passages from unstructured content, and custom tools that run user-defined functions or stored procedures. Web search is one more tool on the agent, and its presence or absence decides whether a Deep Research investigation looks outside the account. Readers comparing this design with other vendors can look at the OpenAI data agent in ChatGPT Work for BigQuery, Snowflake and Databricks, which connects to the same warehouses from the assistant side.

··········

ENABLING WEB SEARCH FOR CORTEX AGENTS IN SNOWFLAKE.

Web search is switched on in two places. First, a user with the ACCOUNTADMIN role enables it for the account in Snowsight under AI & ML, Agents, Settings, by turning on the Web search toggle. Snowflake's documentation describes this step only in Snowsight and does not list an account parameter or SQL statement for it. Second, the tool is added to each agent that should use it, either by editing the agent in Snowsight and turning on Web search under Tools, or through the agents REST endpoint by adding a tool of type web_search named Web Search to the agent's tools array. The documentation does not describe a SQL method for adding this tool.

........

Item	What Snowflake documents
Account switch	ACCOUNTADMIN, Snowsight, AI & ML » Agents » Settings
Agent tool	Snowsight Tools tab or REST API, type web_search
Search provider	Brave Web Search API
Data retention	Zero data retention agreed with Brave
Deep Research duration	Up to 10 minutes per investigation
Release status	Not stated in the release note
Web search pricing	Not stated on the documentation pages

........

Because the setting lives on the agent, administrators can keep web research away from agents that serve sensitive domains and allow it only on agents built for market or competitor questions. That separation is the main governance lever available, since the release note ties web research in Deep Research directly to the agent's web search configuration.

··········

DATA HANDLING WHEN COWORK RESEARCHES THE WEB.

Snowflake states that Cortex Agents use the Brave Web Search API and results from Brave Search's independent web index. The agent builds a search query from the user's input and the relevant conversation context, then merges useful results with the output of its other tools. The documentation is explicit that the query and the results leave Snowflake and traverse the public internet. Snowflake has enabled zero data retention with Brave, under which Brave does not store the search query text, the returned results or any metadata associated with the request.

For Deep Research specifically, the release note adds that the web research has no access to your data. Even so, a query composed from the user's question and context can carry business terms, product names or project names into the outbound request, and security teams should review that behavior against their policies before enabling the account switch. Agents that answer questions about regulated data are reasonable candidates for keeping web search off. The rest of CoWork inherits existing Snowflake controls, including row access policies and column-level security, which Snowflake lists among the product's governance features. Similar questions about semantic layers and governed access arise in Databricks metric view relationships in Unity Catalog.

··········

CONTROLLING COWORK SPEND WITH PER-USER QUOTAS.

Longer investigations that run sub-investigations in parallel and search the web can consume more AI credits than a quick chat turn, and Snowflake's documentation pages for CoWork and web search do not list rates for either. The cost controls are documented, though. Per-user quotas apply the same limit to every user in scope, evaluate each user independently and cover AI domains measured in AI credits, including SNOWFLAKE INTELLIGENCE, the domain name used for CoWork, as well as CORTEX AGENT and AI FUNCTION. Cycles can be monthly, weekly or daily and reset at 00:00 UTC. Blocking is built in for AI domains, and Snowflake warns that enforcement is evaluated within minutes of a spend event, so a user can overshoot the limit slightly, and more so with a large single request.

The syntax below follows Snowflake's per-user quota documentation. The domain value and the limits are illustrative choices for a CoWork rollout.

CREATE SNOWFLAKE.CORE.QUOTA my_quota();
CALL my_quota!ADD_SHARED_RESOURCE('SNOWFLAKE INTELLIGENCE');
CALL my_quota!SET_PER_USER_LIMIT(120);
CALL my_quota!SET_PER_USER_LIMIT(10, 'DAILY');
CALL my_quota!SET_BLOCK_ENFORCEMENT_ENABLED(TRUE, TRUE);

On October 28, 2026, Snowflake made the MY_BUDGET_SUMMARY table function generally available. Any user can call it without extra grants to see the quotas that apply to them, with usage, limit, percentage consumed, cycle dates and block status in one row per quota and cycle. Analysts who run Deep Research regularly can check their remaining allowance before starting a long investigation.

SELECT * FROM TABLE(SNOWFLAKE.LOCAL.MY_BUDGET_SUMMARY());

The illustrative figures below show how the two limits in the example interact for a team of 40 CoWork users. They are not Snowflake defaults or recommendations.

........

Illustrative setting	Value
CoWork users in scope	40
Monthly limit per user	120 AI credits
Daily limit per user	10 AI credits
Maximum monthly exposure before overshoot	4,800 AI credits
Minimum days for one user to reach the monthly limit	12
80% early-warning point per user	96 AI credits

........

For a cap on the whole service, CoWork also supports account-level resource budgets. The CoWork object is tagged, the tag is associated with a budget, and stored procedures can send an alert at a threshold such as 80 percent and revoke access through a dedicated role at 100 percent. Snowflake notes that budget enforcement runs periodically, up to eight hours after a breach, or about two hours with low-latency budgets, so an early alert threshold is advisable.

··········

WHAT DATA TEAMS SHOULD CHECK BEFORE ENABLING IT.

The first check concerns the agents themselves. Deep Research can only reach the semantic views, Cortex Search services and custom tools configured on the agent, so the quality of a report depends on those definitions more than on the web results. Teams that already maintain verified answers and well-described semantic views start from a stronger base, and teams that do not should address that before adding external sources. Google takes a comparable approach with semantic definitions in conversational analytics in Connected Sheets for BigQuery.

The second check concerns review. Because a report now mixes internal findings with web citations, reviewers should confirm which statements come from the warehouse and which come from public pages before a report is forwarded. The citations make this possible, but a web source cited in a report is still a third-party page whose accuracy Snowflake does not vouch for in its documentation.

The third check concerns cost and access: the account switch, the agents with web search enabled, the quotas covering the SNOWFLAKE INTELLIGENCE domain and the resource budget on the CoWork object should all be in place before the feature is announced to business users.

··········

SNOWFLAKE COWORK WEB RESEARCH: CONFIRMED FACTS AND OPEN POINTS.

Snowflake confirms that, from October 30, 2026, Deep Research in CoWork researches the web when the agent has web search enabled, cites web findings next to findings from your data, and keeps the web research without access to your data. Web search runs on the Brave Web Search API with zero data retention, requires an ACCOUNTADMIN to enable it for the account, and is added per agent in Snowsight or through the REST API.

What depends on configuration is the scope of each investigation: the agent's tools, its semantic views and search services, its web search setting and the quotas and budgets applied to its users. What Snowflake has not disclosed on these pages is the release status of the web research capability, the AI credit consumption of a Deep Research investigation and any separate charge for web search queries, so those points should be confirmed with Snowflake or measured in the account's usage views during a pilot.

··········

·····

DATA STUDIOS

·····

datastudios.org