Coding Agents: Model Harnesses and Cortex Code

Two coding agents can use the same model and still produce very different results. The difference is often the harness around the model: the context it receives, the tools it can call, and the instructions that guide its work. This session is for Snowflake developers, data engineers, and architects who want to understand what Cortex Code sends to the model and why that matters.

We will break down the first Cortex Code model call into its actual components, then show why curated active context matters as context windows grow. We will connect those concepts to a matched Snowflake dbt benchmark and inspect the different tool-call patterns observed across coding-agent harnesses.

In this Hands-on Lab, attendees will work through the evidence behind how Cortex Code operates. We will inspect the prompt, system instructions, tool definitions, skill catalog, and selected context included in the first model call; examine how retrieval accuracy changes as context grows; compare real Cortex Code and general coding-agent tool calls from the same Snowflake dbt task; and trace a reported successful file write that persisted malformed content before the agent repaired it.

Attendees will leave with a practical framework for evaluating coding agents beyond the model name alone. They will understand how context selection, tool contracts, and validation affect execution, and how Cortex Code's Snowflake-specific harness changes the path from a user request to completed work.