Snowflake CoCo gets autonomous Cloud Agents and scheduled automations

Cloud Agents and a full-screen CoCo mode

Cloud Agents are the Snowflake managed runtime behind CoCo in the Snowsight interface. They execute code in a secure sandbox and can now run long sessions on their own, according to Snowflake.

The update also adds CoCo mode, a new full-screen experience in Snowsight built for this kind of work. Snowflake lists several supporting features:

• Subagents for multistep tasks.
• Live preview, which shows an app or report while it is being built.
• Native Git integration, which lets CoCo clone, branch, commit and push across repos, then open a pull request for the team to review.

placeholder

Automations, context and shared workflows

Automations give CoCo a recurring job to run inside Snowflake. A user describes the task once, for example checking last night's failed tasks, finding the root cause of each and summarizing what needs a fix. It then runs on a schedule or when an event or webhook starts it. Snowflake says it uses skills and MCP connections, and reports back.

Each automation runs as the user who created it, so existing grants and policies apply to the agent. Every run leaves an audit trail of what CoCo checked, found and did.

Several features aim to give the agent knowledge of the company environment:

• Cortex AI Gateway Tools (public preview) let admins manage MCP connectors, including custom ones, in one place. The same connections work in sessions and automations.
• Skills and plugin sharing, now generally available, let teams share knowledge privately with specific roles through a link, or publish it in the Snowsight catalog, where admins certify the versions that become the standard.
• Agent Profiles define an agent's default model, skills, tools and context, so each new session starts from that setup. Snowflake says this keeps standards consistent and uses fewer tokens.
• Memory, now in public preview across Snowsight, CLI and CoCo Desktop, saves corrections and decisions and brings them back in later sessions. Memories are scoped to the project they came from, so context does not leak between projects.

Model tiers and access from anywhere

Snowflake argues that teams usually pick one model for everything, which means overpaying for routine work or failing on hard tasks. CoCo now offers automatic model tiers. Auto-intelligent keeps every step on the most capable models the company has approved. Auto-efficient uses dynamic model routing, now in public preview, to send routine steps to lighter models and harder ones to a more capable model. Admins set the account default, routing never leaves approved models, and usage shows which model served each request.

CoCo can be used in Snowsight, CoCo Desktop or the CLI, all against the same account and governed data. In the Snowflake mobile app, users can start and steer cloud sessions, approve tool calls and plans, and send follow-ups. Snowflake in Slack brings CoCo into team channels, with answers that follow Snowflake permissions.

The Cortex Agent SDK, now generally available, provides typed Python and TypeScript clients for running CoCo sessions and Cortex Agents from custom code. Users can stream results, run background jobs and manage threads and feedback. The packages are snowflake-cortex-agent-sdk on PyPI and @snowflake/cortex-agent-sdk on npm, and the Python package includes sync and async clients.

What it means for data teams

Snowflake pitches the release as a way to move backlog projects forward while people are away. Its suggested workflow is to package steps as a skill, give the agent a profile, hand over the project, set up an automation to check results each morning, and approve the final step from a phone.

The emphasis throughout is governance: automations run under the creator's permissions, routing stays within approved models, and admins certify shared workflows. For organizations already on Snowflake, the company says this lets teams delegate more work without leaving the controls they already trust. Some of the pieces, including Cortex AI Gateway Tools, Memory and dynamic model routing, are still in public preview.