Snowflake announces a series of updates for CoWork (formerly Snowflake Intelligence). The personal AI agent for knowledge workers is gaining new features, including Cortex Sense, Deep Research, and User Memory. In addition, Snowflake is introducing Cortex Training for fine-tuning custom AI models on company data.
CoWork is evolving into what Snowflake describes as the personal agent for knowledge workers, bringing together data, context, and actions in a single environment. “CoWork gives every employee a personal agent to work smarter across their business data and tools, so they can get insights, take action, and collaborate in one place, all within a secure, governed environment they trust,” said Christian Kleinerman, EVP of Product at Snowflake.
With CoWork, Snowflake introduces the Cortex Sense layer, which automatically brings together business data, business definitions, and operational knowledge for AI agents. Cortex Sense works for both CoWork and CoCo, Snowflake’s coding agent, and provides prebuilt plugins for roles such as finance and sales. This allows companies to deploy production-ready agents without having to rebuild that context from scratch every time.
For complex analyses, Deep Research offers multi-step reasoning across both structured and unstructured business data. Snowflake claims that the underlying agent swarm architecture outperforms single-agent systems by more than a third on its own Hybrid Deep Research Benchmark.
In addition, CoWork continuously learns from individual user behavior via User Memory. Users can schedule tasks and receive role-specific actions via tools like Gmail or Slack. New features also include an iOS app for CoWork, a Slackbot, and a Microsoft Excel extension.
Cortex Training: Train your own models where the data already resides
In addition to the CoWork updates, Snowflake is launching Cortex Training. This allows organizations to fine-tune open-weight foundation models, such as those from the Qwen or Mistral families, on their own data, without managing GPU infrastructure or moving data to external environments. According to initial customer reviews, it is possible to run up to twice as many training runs for the same GPU budget.
Snowflake also gives customers more choice in AI models through Cortex AI. In addition to existing providers such as Anthropic, OpenAI, Google, Meta, Mistral, and DeepSeek, Snowflake is now adding models from SpaceXAI. Over 13,900 customers already trust Snowflake with their data.