Building AI Agents with Cortex Code and CoWork

Snowflake for Developers/Guides/Building AI Agents with Cortex Code and CoWork

Quickstart

Building AI Agents with Cortex Code and CoWork

Carlos Serrano

Overview

This guide walks you through building a complete AI-powered analytics agent using Snowflake Cortex Code (CoCo) — the AI-native IDE for Snowflake. You will go from raw data to a production-ready Cortex Agent with evaluations, all driven by natural language prompts.

The lab uses a convenience store hot food sales dataset (100 stores, 100 items, 539K transactions) and demonstrates how Cortex Code automates the creation of semantic views, agents, skills, and evaluation frameworks.

┌─────────────────────────────────────────────────────────┐
│                  HOL_COCO_COWORK Database                 │
├───────────────┬──────────────────┬──────────────────────┤
│  DATA Schema  │  TOOLS Schema    │  AGENTS Schema       │
│               │                  │                      │
│  DIM_STORE    │  Semantic View:  │  Cortex Agent:       │
│  DIM_ITEM     │  HOT_FOOD_SALES  │  HOT_FOOD_SALES     │
│  FACT_ITEM    │  _ANALYTICS      │  _AGENT             │
│  _SALES       │                  │                      │
│               │  (8 VQRs,        │  (text-to-SQL tool,  │
│  100 stores   │   2 relationships│   2 skills,          │
│  100 items    │   3 tables)      │   orchestration +    │
│  539K sales   │                  │   response instrs)   │
│               │  Skills Stage:   │                      │
│               │  anomaly_detect  │                      │
│               │  sales_report    │                      │
└───────────────┴──────────────────┴──────────────────────┘

Prerequisites

A Snowflake account with ACCOUNTADMIN access
Cortex Code desktop application installed and connected to your Snowflake account
Access to a Snowflake Workspace

What You Will Learn

How to use Cortex Code to build Snowflake objects from natural language prompts
How to create semantic views with verified queries for Cortex Analyst
How to build a Cortex Agent with orchestration instructions and server-side skills
How to evaluate agent performance using Snowflake's built-in evaluation framework

What You Will Build

A star-schema data model for convenience store hot food sales
A semantic view with 8 verified queries (VQRs)
A Cortex Agent with anomaly detection and sales report skills
An evaluation framework measuring answer correctness and logical consistency

Setup Environment

This step creates all the infrastructure needed for the lab. You will run SQL and Python scripts in a Snowflake Workspace.

Required Files

Download the Setup folder which contains all scripts and data needed and extract them:

File	Type	Purpose
01_setup.sql	SQL	Creates warehouse, database, schemas, stages, and tables
02_copy_files.py	Python	Uploads CSV data and skill files to stages
03_load_data.sql	SQL	Loads data into tables and verifies row counts
data/dim_store.csv	CSV	100 convenience stores
data/dim_item.csv	CSV	100 hot food items
data/fact_item_sales.csv	CSV	539K sales transactions

Download the Skills and Prompts folder and extract them:

Step 1: Load Project Files into a Workspace

Download all the files listed in the Required Files table above (or clone the companion repository)
In Snowsight > Projects > Workspaces, click + Workspace to create a new blank workspace
Use the Upload button (or drag and drop) to upload the following folder structure:
- Setup/ — including the data/ subfolder with all three CSV files
- Skills/ — both skill subfolders with their SKILL.md files
- Prompts/ — all three prompt files (Optional)
Verify all files appear in the workspace file browser before proceeding

Step 2: Create Infrastructure (SQL)

Open Setup/01_setup.sql in the Workspace and click Run All. This creates:

-- Warehouse and compute pool
CREATE WAREHOUSE IF NOT EXISTS HOL_WH
  WAREHOUSE_SIZE = 'XSMALL'
  AUTO_SUSPEND = 60
  AUTO_RESUME = TRUE;

CREATE COMPUTE POOL IF NOT EXISTS HOL_COMPUTE_POOL
  MIN_NODES = 1
  MAX_NODES = 1
  INSTANCE_FAMILY = CPU_X64_XS
  AUTO_RESUME = TRUE
  AUTO_SUSPEND_SECS = 300;

-- Database and schemas
CREATE DATABASE IF NOT EXISTS HOL_COCO_COWORK;
CREATE SCHEMA IF NOT EXISTS DATA;
CREATE SCHEMA IF NOT EXISTS TOOLS;
CREATE SCHEMA IF NOT EXISTS AGENTS;

-- Tables: DIM_STORE, DIM_ITEM, FACT_ITEM_SALES
-- File format, stages, and foreign key constraints

Step 3: Upload Files to Stages (Python)

Open Setup/02_copy_files.py in the Workspace and click Run All. This uploads:

CSV data files to @HOL_STAGE
Agent skill files to @SKILLS_STAGE

Step 4: Load Data (SQL)

Open Setup/03_load_data.sql in the Workspace and click Run All. This loads data into all three tables.

Verify the final query output shows:

TABLE_NAME	ROW_COUNT
DIM_STORE	100
DIM_ITEM	100
FACT_ITEM_SALES	539,215

NOTE: If row counts don't match, re-run the scripts in order. The fact table uses a staging table with JOINs to resolve UUID foreign keys.

Create Semantic View

From this point forward, all steps are performed in Cortex Code Desktop (CoCo). Open the project folder in CoCo and use the chat panel.

A semantic view is an AI-ready data layer that maps business concepts to your tables. It enables Cortex Analyst to generate accurate SQL from natural language questions.

Run the Prompt

In the Cortex Code chat panel, type:

/semantic-view @semantic_view.md

The @ symbol attaches the file Prompts/semantic_view.md as context for the command.

What CoCo Does

Cortex Code will automatically:

Discover all tables in HOL_COCO_COWORK.DATA
Generate a semantic model with dimensions, facts, and relationships
Create 8 verified queries (VQRs) covering common sales analytics questions
Validate the YAML against Snowflake
Deploy the semantic view to HOL_COCO_COWORK.TOOLS

Verified Queries Included

The semantic view includes verified queries for:

Total revenue
Total revenue by category
Total revenue by state
Top 10 stores by revenue
Top 10 best-selling items by quantity
Monthly revenue trend
Average discount percentage by category
Total transaction count

Expected Result

Semantic view HOL_COCO_COWORK.TOOLS.HOT_FOOD_SALES_ANALYTICS is created with:

3 tables (FACT_ITEM_SALES, DIM_STORE, DIM_ITEM)
2 relationships (fact to each dimension)
8 verified queries

Create Cortex Agent

A Cortex Agent is an intelligent entity that reasons over your data using tools (text-to-SQL, search, skills) and responds to user questions conversationally.

Run the Prompt

In the Cortex Code chat panel, type:

/cortex-agent @agent.md

What CoCo Does

Cortex Code will:

Create a workspace directory for the agent configuration
Build the agent specification with:
- Orchestration instructions — role context, tool selection logic, boundaries, and business rules
- Response instructions — assertive tone, chart generation, multilingual support
- Tool configuration — pointing to HOL_COCO_COWORK.TOOLS.HOT_FOOD_SALES_ANALYTICS
- Skills — anomaly detection and sales report generation from @SKILLS_STAGE
Validate and deploy the agent to HOL_COCO_COWORK.AGENTS

Agent Skills

The agent is equipped with two server-side skills:

Anomaly Detection — Performs z-score analysis over a 7-day rolling window to find unusual spikes or drops in revenue, quantity, or transactions. Supports grouping by store, category, state, or item.

Sales Report Generator — Produces structured executive reports with summary metrics, top products, monthly trends, missed opportunities, and recommended actions for a specific store or state.

Expected Result

Agent HOL_COCO_COWORK.AGENTS.HOT_FOOD_SALES_AGENT is created and ready to answer questions.

Run Evaluations

Evaluations measure how well your agent answers questions against ground truth data. This step creates a test dataset and runs automated scoring.

Run the Prompt

In the Cortex Code chat panel, type:

@evaluations.md

What CoCo Does

Cortex Code will:

Query the underlying tables to compute ground truth answers
Create an evaluation dataset with 10 questions across 5 categories:
- Basic metrics — total revenue, transaction count, average transaction value
- Dimensional analysis — revenue by state, by category, by discount
- Trend analysis — monthly revenue patterns
- Rankings — top store, top item
- Filter analysis — spicy items performance
Register the dataset using SYSTEM$CREATE_EVALUATION_DATASET
Run the evaluation measuring:
- answer_correctness — factual accuracy of the agent's responses
- logical_consistency — coherence and reasoning quality
Present results with per-question scores

Expected Result

Evaluation scores of approximately:

Answer Correctness: ~93%
Logical Consistency: 100%

Test the Agent in CoWork

Now that the agent is deployed, you can interact with it through Snowflake CoWork.

Access CoWork

Open Snowflake CoWork and ensure:

Your role is set to ACCOUNTADMIN
Your warehouse is set to HOL_WH
Your agent is set to HOT_FOOD_SALES_AGENT

Sample Questions

Try these questions to test the agent's capabilities:

Basic Analytics

What is the total revenue for Q1 2025?
How many transactions happened in March?

Dimensional Analysis

Which state generates the most revenue?
What are the top 5 categories by sales volume?

Trend Analysis

Show me the monthly revenue trend with a chart.
How did February compare to January?

Anomaly Detection (Skill)

Are there any unusual revenue patterns by category?
Detect anomalies in transactions by store.

Sales Reports (Skill)

Generate a sales report for QuickStop #001.
Give me an executive summary for the state of NY.

Conclusion And Resources

Congratulations! You have built a complete AI-powered analytics agent using Cortex Code — from raw data to a production-ready Cortex Agent with automated evaluations.

What You Learned

How to set up a star-schema data model in Snowflake for analytics
How to use Cortex Code's /semantic-view command to auto-generate semantic views with verified queries
How to use Cortex Code's /cortex-agent command to build agents with orchestration instructions, tools, and skills
How to create ground truth evaluation datasets and run automated agent evaluations
How to interact with your agent through Snowflake CoWork

Key Takeaway

Steps 2-4 were entirely driven by single natural language prompts in Cortex Code. The AI-assisted development workflow lets you build production Snowflake objects from high-level instructions — no manual YAML authoring, no SQL debugging, no configuration files.

Related Resources

Updated 2026-07-07

This content is provided as is, and is not maintained on an ongoing basis. It may be out of date with current Snowflake instances