CoCo: Snowflake's AI Coding Agent for Data Engineers

Details

Snowflake has published a detailed guide to CoCo, its AI coding agent designed specifically for data engineers working within the Snowflake ecosystem. Unlike generic coding agents, CoCo is trained on Snowflake-specific context and extended through Skills (custom instructions) and Plugins (tool integrations) that let teams encode institutional knowledge into reproducible workflows.\n\nThe post identifies four core best practices: start with minimal prompts and iterate from failures rather than overengineering upfront; trust the model's existing intelligence rather than over-explaining; treat the context window as a finite, shared resource where every sentence must earn its place; and prioritize reproducibility over novelty.\n\nTwo hard limits stand out. First, AI agents should not replace enterprise data tooling — dbt handles transformations, and DCM tools like schemachange, Flyway, or Terraform handle deployments. Agents help design and build pipelines but should not execute in production. Second, because agent outputs are nondeterministic, allowing them to make changes directly in production introduces fragility that engineering discipline alone cannot contain.\n\nThe post comes from Snowflake's own blog, so it functions partly as product marketing. That said, the best practices are substantive and reflect genuine challenges practitioners face when integrating AI agents into data workflows at scale.