Snowflake is introducing Cortex Agents through a public preview. This will allow customers to attach Agents to their data. Furthermore, the company is embracing Anthropic with its Claude-LLM.
No AI without agents, if the big technology companies are to be believed. Consequently, Snowflake also wants to give its customers access to AI agents and is doing so through Cortex Agents. Cortex is the name for Snowflake’s AI platform within the company’s data cloud solution, and agents click into it.
Cortex Agents is available as a public preview. The agents can extract insights from structured and unstructured datasets within the data cloud. They are built on the Cortex AI foundation, including the retreival capabilities of that platform. Snowflake sees agents as tools that, based on a question, search for answers in relevant datasets on their own, possibly by combining them autonomously.
Combined workflow
Cortex Agents are built on Cortex Analyst and Cortex Search. Analyst generates structured SQL, and Search specializes in unstructured data. Snowflake describes the capabilities of an agent in action as follows:
- The agent can analyze a broad query from a user and split it into subtasks, with one task covering, for example, structured data from an order system, and others from unstructured data from product descriptions.
- The agent will then create a plan and retrieve the relevant structured and unstructured data from the relevant systems.
- This is followed by a moment of reflection: the agent evaluates the result and chooses the next step. This may be a secondary question, additional steps or a final answer. Reflection ensures accuracy.
Cortex Agents is new but also not. In fact, Snowflake announced agents last year, but then under the name Snowflake Intelligence. It makes more sense to put the functionality along under the Cortex umbrella.
The core of Snowflake is a zero copy data warehouse in the cloud where all customer data is securely stored. On top of that is a strong security and policy layer with access rights. The AI agents adhere to those policies.
Claude
To help power the agents, as well as provide more capabilities for other Cortex capabilities, Snowflake is now integrating Anthropic’s Claude LLM. Claude can handle structured and unstructured data, including generating SQL tasks.
Snowflake wants to remain vendor-agnostic within its platform. Customers can unleash AI on their data through Cortex, but have free choice about which models they use to do so. For certain targeted applications, the company provides its own solutions; in other cases, customers can choose. Claude from Anthropic is an interesting new option in this regard. Just recently, Snowflake announced that DeepSeek-R1 is also available within the platform.