Snowflake Launches Unattended CoCo AI Automations in Public Preview
Snowflake opened CoCo automations to public preview, allowing recurring unattended AI-agent runs in Snowflake-managed sandboxes with auditable Cortex threads.
Why it matters: CoCo shifts Snowflake AI from interactive assistance toward unattended execution, making identity, permissions, MCP access and auditability central enterprise concerns.
Snowflake has opened CoCo automations to public preview, allowing enterprise AI agents to execute recurring tasks inside Snowflake without a person remaining online.
Users can schedule prompts to run inside Snowflake-managed sandboxes for workflows including anomaly detection, data-quality checks, warehouse-cost monitoring, repository maintenance and recurring reports.
Every automation run creates a Cortex thread containing the agent's messages, tool calls, results and final response, providing an auditable record that teams can inspect after execution.
Automations can also connect to MCP servers, Snowflake stages and secrets, giving agents access to external tools and persistent resources.
The feature introduces important governance considerations. Snowflake documents that the account-level EXECUTE AGENT TASK privilege is granted to the PUBLIC role by default unless administrators revoke it and assign the capability to selected roles.
Scheduled runs execute as the user who created the automation, while unattended execution means there is no person present to approve each tool action interactively. Snowflake recommends applying least privilege, attaching only required tools and secrets, and avoiding broad prompts that could act on unintended objects.
CoCo automations are another sign that AI agents are becoming a native execution layer inside enterprise data platforms rather than merely an interface sitting on top of them.