The Tool Desk
Outbyte Driver Updater FREEFix the driver behind crashes, sound loss and screen glitchesFind Drivers →Outbyte PC Repair FREERepair Windows errors before they cause bigger problemsFix Now →Some links on this page are affiliate links: if you buy through them we may earn a commission, at no extra cost to you.
Snowflake’s February 12, 2025 announcement introduced Cortex Agents in public preview as an API-first way to build enterprise agents that combine structured warehouse data with unstructured documents. The original design coordinated Cortex Analyst for natural-language SQL and Cortex Search for document retrieval, using an LLM to choose the appropriate tool. By August 2026, Snowflake’s documentation described a substantially broader managed runtime with persistent threads, code execution, charts, custom tools, MCP connectors, evaluations and monitoring.
That distinction matters for anyone budgeting or designing an enterprise AI system: the launch was a preview, not a finished chatbot subscription, and current capabilities, model availability, pricing and production guidance are different from the 2025 announcement.
What Snowflake announced on February 12, 2025
Snowflake announced the public preview of Cortex Agents, positioning APIs—not just a Snowflake-hosted chat experience—as the way developers could embed data agents in their own applications. The announcement described agents that could coordinate large language models with Cortex Analyst and Cortex Search.
Analyst translates questions about governed, structured data into SQL through a semantic model. Search retrieves relevant passages from indexed documents and other unstructured sources. Cortex Agents supplied the orchestration layer, deciding which capability to use and combining the results. Snowflake’s announcement also highlighted its Anthropic partnership and Claude 3.5 Sonnet; that model reference is historical, not a description of today’s model catalog.
#1 Best Overall
This was different from Snowflake Intelligence, the business-user-facing experience Snowflake had positioned as a low-code interface. Cortex Agents were the developer-facing runtime. Cortex Analyst and Cortex Search were specialist tools inside that runtime, rather than interchangeable names for the same product.
Snowflake’s thesis was that useful enterprise AI depends on accessible, well-modeled and governed data. That is a reasonable architecture argument, but it is not a guarantee of accurate answers: results still depend on semantic definitions, source quality, retrieval, permissions and model behavior.
How a Cortex Agent works
Current documentation describes a managed loop:
- Plan: parse the request, resolve ambiguity, split complex work into subtasks and select tools.
- Use tools: call Analyst, Search, code execution or another configured tool.
- Reflect and respond: inspect the results, decide whether another call is needed, then produce an answer and any available citations or visual output.
A typical request might be: “Which customers whose contracts expire in the next 90 days generated less revenue this quarter, and what renewal risks are mentioned in their account notes?” Analyst needs semantic views defining customers, revenue, quarter and contract expiry. Search needs an indexed collection of account notes or PDFs. The agent then has to reconcile the returned customer identifiers and evidence while honoring the caller’s data permissions.
Windows Errors? Fix Them Before They Spread
Repair common Windows errors and clear accumulated junk for a smoother, more stable PC - no reinstall needed.Free scan · no reinstallCrashes, No Sound, or Screen Glitches?
Random freezes, missing sound and display glitches usually trace back to one bad driver. Find and replace yours safely.Free scan · under a minuteThat is more than a fixed prompt or basic retrieval-augmented generation pipeline. A conventional RAG bot generally retrieves documents and drafts an answer. Cortex Agents can select among SQL, search, Python, charts, stored procedures, remote tools and other agents, while maintaining state in persistent threads. The managed service reduces the infrastructure a team must operate for planning, routing, state and run telemetry, but it does not remove the need for testing and controls.
What the current platform includes
As documented in August 2026, Cortex Agents can include:
- Cortex Analyst: natural-language questions over structured data through semantic views.
- Cortex Search: retrieval over unstructured content with adjustable search parameters.
- Code execution: Python in an isolated sandbox.
- Data to Chart: visualizations based on tool results.
- Custom tools: stored procedures and user-defined functions for business logic or backend calls.
- Agent skills: reusable instruction-and-script bundles.
- MCP connectors: remote tools such as Jira, Salesforce or customer applications.
- Agent toolsets: references to tools exposed by other agents.
- Web search: public-web retrieval when enabled for the account.
Agents can be created in Snowsight, with SQL or through APIs, then called with the agent:run operation over REST. Run events expose tool calls and execution details. Threads preserve conversational context across requests. Snowflake also documents monitoring, traces, feedback and evaluations so teams can examine tool selection and answer quality rather than treating the agent as an opaque chat box.
Implementation prerequisites
A production design normally needs all of the following:
- A Snowflake account with the relevant Cortex capabilities and an available model in the account’s cloud and region.
- Semantic views with explicit business definitions for Analyst.
- Search services, indexed content, metadata and a refresh policy for documents.
- An agent definition containing a model, instructions and approved tools.
- Roles and object privileges for the agent and every object its tools can access.
- Secure API authentication and authorization-token handling.
- Warehouses or other compute resources for custom tools and warehouse-backed work.
- Evaluation, human-review and cost-monitoring procedures before user exposure.
Calling an agent requires the SNOWFLAKE.CORTEX_USER or SNOWFLAKE.CORTEX_AGENT_USER database role, privileges on the agent object and privileges on objects used by configured tools. Snowflake’s documentation should be treated as the authority for the exact account configuration and API reference; the launch article does not provide enough information to safely copy a complete production request.
Rank #3
Model availability varies by cloud, geography and the CORTEX_ENABLED_CROSS_REGION setting. The current catalog includes Anthropic, OpenAI and Google families, but not every model is available everywhere, and public-preview models are not intended for production workloads. Cortex Agents APIs also are not supported from a Streamlit in Snowflake app using a warehouse runtime; that use case requires a container runtime.
Governance and security questions
Snowflake-native governance is the strongest reason to consider the service, but it must be designed rather than assumed. Use least-privilege roles, narrow semantic views, filtered search services and tool-specific permissions. Test whether a user who can call an agent can see only the rows, documents and actions intended for that user.
Custom procedures, UDFs and MCP connectors can expand an agent’s reach into operational systems. Each connector adds identity, secrets, network, logging and third-party-risk questions. Web search creates an external-data path that may be unsuitable for regulated workloads. Documents and tool responses can also contain malicious or misleading instructions, so treat retrieved text as untrusted input and validate consequential actions.
Do these 3 things before closing this tab:
1Clear out junk files and repair common Windows errors2Scan for outdated or missing drivers - takes under a minute3Repair Windows errors before they cause bigger problemsSnowflake warns that agent responses and citations are not guaranteed to be accurate and should be reviewed before being served to users. Measure at least tool-selection accuracy, answer correctness, citation quality, latency, cost per successful task and data-access violations. Include adversarial tests, stale-document tests and ambiguous business questions—not only happy-path demos.
Rank #4
Preview status and operational limitations
The February 2025 service was a public preview. Snowflake’s preview policy says preview features may change, may have defects or incomplete corner-case handling, and should primarily be used for evaluation and testing rather than production systems or production data. That warning should not be omitted from a retrospective launch story.
Common failure modes remain:
- Bad semantic views produce SQL that is syntactically valid but uses the wrong metric or time period.
- Incomplete, poorly chunked or stale documents weaken Search results.
- The planner can choose the wrong tool or combine contradictory results.
- Multi-step runs increase latency and consumption, especially when code, Analyst and Search are all invoked.
- Generated explanations and citations can be incomplete or wrong.
- Broad custom tools and MCP access increase the security and audit burden.
How pricing works
Cortex Agents is not presented as a simple per-seat chatbot subscription. Snowflake documents several consumption components: orchestration token usage, Cortex Analyst token usage, Cortex Search charges based on index size and persistence, and warehouse costs for custom tools. Code execution, concurrency, monitoring and repeated retries can also affect the workload’s total spend.
There is no reliable flat per-agent or per-request price in the supplied documentation. Build a workload estimate from questions per day, average turns, model choice, Analyst and Search calls, index size and retention, warehouse size and runtime. Budget by completed successful task, not merely by user or conversation, and verify the current service-consumption documentation with Snowflake before signing a commitment.
How it compares with alternatives
| Option | Most natural fit | Main trade-off |
|---|---|---|
| Snowflake Cortex Agents | Snowflake-centered governed SQL plus document reasoning | Consumption pricing, regional limits and Snowflake coupling |
| Databricks Mosaic AI | Databricks lakehouse, Unity Catalog and ML estates | Less attractive when governance and data are already Snowflake-native |
| Microsoft Fabric | Microsoft 365, Power BI, Azure and Fabric users | Weaker fit without meaningful Microsoft adoption |
| Vertex AI Agent Builder or Amazon Bedrock Agents | Cloud-native application and API integration | Requires rebuilding or exposing Snowflake-specific semantic governance |
| LangGraph or LlamaIndex | Teams prioritizing portability and orchestration control | You must assemble identity, state, tools, sandboxing, evaluation and observability |
Verdict
Cortex Agents is most compelling when an organization already governs important structured and unstructured data in Snowflake and wants a managed, multi-step runtime rather than another orchestration stack. It is less compelling for simple document search, highly deterministic workflows, data estates primarily outside Snowflake, strict latency or cost ceilings, or teams whose strategy requires self-hosted portability.
Best Value
The 2025 announcement was significant because it exposed Snowflake’s data-agent idea through APIs. The current product is broader, but its success still depends on the unglamorous foundations: semantic modeling, indexing, least privilege, evaluation and cost controls.
Frequently Asked Questions
Is the February 2025 Cortex Agents announcement still an accurate description of the product?
No. It accurately describes the original public preview, but current Cortex Agents documentation covers a broader managed platform with additional tools, threads, monitoring, evaluations and integrations.
Does Cortex Agents guarantee accurate answers or citations?
No. Snowflake says responses and citations require review. Accuracy depends on semantic views, retrieval quality, source data, permissions, tool choice and model behavior.
Quick wins for a faster PC:
Scan for outdated or missing drivers - takes under a minuteDriver Scan →Repair Windows errors before they cause bigger problemsFix Now →Can Cortex Agents be priced like a per-user chatbot?
Not from the available documentation. Costs are consumption-based across orchestration and Analyst tokens, Search index persistence and warehouse-backed custom tools.
Quick Recap
Product prices and availability are accurate as of the date/time indicated and are subject to change. Any price and availability information displayed on Amazon at the time of purchase will apply.
