A month with Cortex Code (CoCo) in Snowflake: what it really changes

The launch of Cortex Code (CoCo) has not gone unnoticed among Snowflake users. It is already shaping up to be a real shift for your environments. Here is our feedback and what you need to know.

Generative AI is already everywhere in the data ecosystem, to the point where it raises the question of whether dashboards are still needed. But here, the promise is different. This is not just an assistant that can write code, it understands your Snowflake environment.

In other words, this is not a generic copilot. It is an agent embedded in your environment. It can navigate your schemas, your objects and your governance, helping you move faster while staying within Snowflake.

So is this just another trend, or a real business accelerator? That is what we are going to clarify.

Cortex Code in 2 minutes

• In Snowsight, CoCo is integrated into a side panel and understands the context of your session. You can ask a question, and it analyzes your open query as well as the objects you are exploring.
• For users working with external IDEs, the CLI version allows you to interact directly with your local files and Git repositories. This moves towards a more agent-based approach to simplify the setup of end-to-end pipelines. It also makes it possible to structure and customize interactions by adding mechanisms such as agent rules.

It is worth noting that Cortex CLI is not currently available on Windows, so you will need to use WSL, Windows Subsystem for Linux, to run it.

What changes here is that the assistant no longer works alongside your platform, but inside it. It relies on the actual metadata of your Snowflake account, stays aligned with your governance and suggests context-aware actions through dedicated skills rather than generic responses.

So Cortex Code is not just about writing code faster. It helps reduce the time between a question, an objective and the right action.

Feedback from a real use case

It all started when I noticed a small but consistent usage on one of my warehouses, around 0.4 credits every day. Nothing critical, but the regularity caught my attention. With everything that has been set up recently on the account, from dynamic tables to dbt projects, as well as agents and notebooks, the cause could come from multiple sources.

Instead of running a manual investigation, I took the opportunity to ask CoCo for help with a simple question: My GLEM_WH warehouse is running every day when it shouldn’t. Identify what is triggering it and what I can do to stop this credit consumption.

Its approach is methodical. It starts with around ten queries to identify or rule out the most likely causes. It then analyzes the hourly pattern and reaches a clear conclusion: the query is running every hour, day and night. Once the query is isolated, the next step is to trace it back to its source. CoCo then provides a key insight: the same session, executed through Snowflake WebApp, has remained active for several weeks. The issue comes from a human action, and in this case, I am the one responsible !

CoCo’s first hypothesis is that a Snowflake dashboard refresh is scheduled.

I then take over to check the status of my dashboards. Nothing unusual, false alarm.

I continue iterating by feeding this information back to CoCo, which suggests new actions such as killing sessions to better identify the source. Still nothing changes, the queries keep running at fixed times.

Then comes the breakthrough! Based on the insights provided by Cortex Code, I take another look at my dashboards. The dashboard itself is not scheduled, but one of its filters is, and it runs every hour.

The root cause is identified and resolved after just a few minutes of interaction with CoCo. Impressive!

Can everything be delegated to CoCo?

What do I take away from this example? Not every analysis can be solved on its own, even with an assistant like CoCo. However, it is clearly a powerful accelerator to support day-to-day work in Snowflake!

It is also worth noting that even if CoCo has access to documentation, it does not always use it reliably. In another case, when setting up a Streamlit application, it suggested a parameter that does not exist. It corrected itself as soon as I asked it to check the documentation.

Finally, it is important to remember that Cortex Code is not an autopilot. In workspaces, CoCo will ask for validation by default before modifying or updating a query. And this is clearly the right approach, especially in production environments!

Cortex Code is a powerful accelerator, but its effectiveness depends directly on the framework you set for it.

Using Cortex Code: 4 best practices

Any use of AI requires a clear framework. Here are the four key practices I take away for Cortex Code:

1. Communicate naturally and focus on the objective rather than how to achieve it, while iterating to refine and challenge the outputs.
2. Anticipate and understand how the AI behaves by using the /plan feature as soon as a task becomes complex or sensitive.
3. Always validate the code generated by CoCo to fully understand its impact.
4. Ensure security by staying vigilant about the permissions granted.

The key point is the level of customization. CoCo in Snowsight remains a highly effective contextual assistant for exploring, understanding and analyzing. But as soon as you want to introduce more structure, with rules, skills or hooks, the CLI becomes essential.

Cortex Code is a powerful accelerator, but its effectiveness depends directly on the framework you set for it.

CoCo’s strength: reducing the time between idea and action

Cortex Code does not aim to replace data engineers or admins. It offers something more valuable: reducing the time between an idea and execution by leveraging the Snowflake context and its governance.

In other words, it is not just about saving development time, but about reducing the time needed to frame the right approach. And this is often where the real bottlenecks lie.

The real question is not “can Cortex Code write SQL,” but “does it help me decide faster what to build, where to do it and why.” That is what turns an assistant into a true accelerator. And with the CLI now supporting tools like Airflow and dbt, it is set to become essential!