Snowflake CoCo AI coding agent keeps data inside

Snowflake CoCo AI coding agent is positioned as the way data engineers build reproducible pipelines without letting data leave Snowflake's security perimeter.

Jeremiah Hansen, Snowflake sales engineer and author of the September 4 post, frames the pitch around reproducibility over raw code generation.

How CoCo actually runs

CoCo runs in three places that share the same intelligence: a CLI for terminal work, a VS Code-based Desktop, and a no-install version inside Snowsight for long-running cloud tasks.

Inference stays inside Snowflake infrastructure rather than routing to a third-party provider, so schemas, warehouse configs and query history are processed under existing governance and classification policies.

Think of it like a shop tool that automatically pulls the right jig when it recognizes the material: CoCo loads built-in skills for Dynamic Tables, Snowpipe Streaming, Openflow, Spark migration, dbt, Snowpark and DCM the moment your prompt matches the domain.

Why this matters, and what still breaks

For builders the security promise matters because most generic agents require shipping context to external LLMs, which breaks governance for regulated data.

Hansen is blunt about the remaining gap: agents are nondeterministic by design, so letting them write directly to production without dbt, schemachange, Flyway or Terraform creates fragile deployments that no instruction file can fully tame.

Snowflake answers with packaging: skills defined by SKILL.md encode team conventions and scripts, while plugins bundle those skills with subagents, slash commands, hooks and MCP servers into a versioned, validatable unit deployed via its registry with cortex plugin validate, install and update.

Reproducibility is now the job, not prompting.