Snowflake CoCo automations run as you — with your default role

Snowflake put CoCo automations into preview in the CLI and Snowsight on August 21, turning a prompt into a recurring, unattended agent run that fires whether or not your laptop is open. It is a genuinely useful primitive. It also has a permissions detail buried in the docs that will bite anyone who schedules first and reads later.

What actually happened

CoCo is Snowflake's rebranded Cortex Code agent. An automation converts a saved prompt into a scheduled run executed inside Snowflake's managed environment. Per Snowflake's documentation, each one is stored as a Snowflake AGENT TASK in your personal database under USER$.PUBLIC, with a COCO_ROUTINE_ prefix on the internal task name. Schedules support intervals, daily times, and weekly times, with a floor of once per hour. Threads are retained for two months.

The runtime detail worth reading twice: a run executes as the user who created the automation, using that user's default role and default secondary roles — not the role that was active in the CoCo session where the automation was created. The docs say it plainly: "A run does not use the role that was active in your CoCo session when you created the automation."

The sandbox mounts the user's USER$.PUBLIC.DEFAULT$ workspace stage at /workspace, so files persist across runs. There are no event-based triggers yet, only fixed schedules, and you cannot attach to a run in progress or resume one interactively. This is still preview, in supported commercial regions.

Why scheduled agent runs matter for your business

The permission model is the product. Most teams work in Snowflake under an elevated role they switched into for the task at hand, then create an automation from that session and assume it inherits. It does not. Your automation runs on your default role forever after, which means it either quietly fails at 2am or — worse, if your default role is broad — does more than the prompt you tested. Set the creating user's default role deliberately, or create automations under a purpose-built service user.

Unattended is a different risk class than interactive. An agent you watch has you as the error handler. An agent on a cron does not. You cannot attach to a run in progress, so your only observability is the thread afterward and whatever the run wrote. Build the automation to log its own decisions and to fail loudly, because nothing else will.

Recurring work is where agent economics actually close. The pitch for coding agents has mostly been faster feature work, which is hard to measure. Hourly data validation, freshness checks, reconciliation against last night's load, a rebuild that runs before anyone logs in — that is repeatable, boring, and exactly the shape of work worth automating. The value is in the schedule, not the model.

Sources: Snowflake CoCo automations documentation, Snowflake release notes.