What you will be able to do
- Describe how Cortex Agents plan, use tools and reflect, and which tools they combine
- Place Snowflake CoWork, Cortex Code (Snowsight and CLI) and Snowflake Copilot inline by who uses them and where
- Choose among the AI & ML Studio, SQL, Python and REST API interfaces for a scenario
- Explain how to bring your own model with the Model Registry CustomModel class and serve it on Snowpark Container Services
1.Cortex Agents: orchestrating Search, Analyst and other tools
Cortex Agents is Snowflake's fully managed platform for AI agents that combine structured and unstructured data in one governed workflow. An agent writes SQL over structured data through Cortex Analyst semantic views and retrieves from documents through Cortex Search. It can also run Python in an isolated sandbox, generate charts, call custom tools built from stored procedures or UDFs, and reach remote MCP servers or web search. Most tools run entirely inside Snowflake. Web search and MCP connectors are the exception: they send requests over the public internet.
An agent is a reusable schema-level object that bundles a model, its tools and its instructions. Snowflake recommends setting the model to auto. Each request goes through a reasoning loop: the agent plans, calls the tools it chose, then reflects on the results. From there it answers, asks a clarifying question, or calls another tool. Threads keep conversation context between turns, so your client does not have to. Tools are what give an agent access to your data. An agent with no tools can still hold a conversation, but it answers only from the model's general knowledge.
Checkpoint 1 of 7· Put it in order
Put the steps of a Cortex Agent's reasoning loop in order
- 1.Reflect and respond: evaluate tool results and decide whether to answer, clarify or call another tool
- 2.Use tools: call Cortex Analyst, Cortex Search, code execution or other selected tools
- 3.Plan: parse the request, resolve ambiguity, split it into subtasks and choose tools
The agent plans, uses tools, then reflects. The loop repeats as needed within a single request.
“Reflect and respond: The agent evaluates the results from each tool to decide what to do next”Source: docs.snowflake.com
Sources1
2.Snowflake Intelligence and the business-user experience
The exam guide lists Snowflake Intelligence as a feature. The sources for this lesson do not document a product by that name, so this section does not describe it. What the sources do document is a ready-made application for business users in the same suite: Snowflake CoWork. Snowflake's AI overview lists CoWork next to Cortex Agents, Cortex Analyst and Cortex Search. It is described as a ready-to-use agentic app with a conversational interface, where business users ask questions in natural language about both structured and unstructured enterprise data. It uses data agents to understand the question, run the analysis and produce insights. Snowflake's security and governance policies still apply.
The difference from the previous section is who builds what. With Cortex Agents, developers create the agent object and call it from their own app through the REST API. The Agents documentation notes that users can also talk to those same agents directly in Snowflake CoWork and in Cortex Code. No custom front end is needed.
Checkpoint 2 of 7· Check yourself
An operations director wants to ask natural-language questions across sales tables and policy documents in a ready-made Snowflake app, without a developer building a chat front end. Which documented capability fits?
CoWork is the ready-to-use conversational application for business users. The Cortex Analyst API needs a team to build the client, and Cortex Code is a developer tool.
“Snowflake CoWork is a ready-to-use agentic application with an intuitive, conversational interface that helps business users discover and act on deep insights.”Source: docs.snowflake.com
3.Cortex Code in Snowsight, the Cortex Code CLI and Snowflake Copilot inline
Snowflake's AI overview lists Cortex Code, Cortex Code in Snowsight and Cortex Code CLI as part of the Cortex suite. The product documentation calls it Snowflake CoCo. It is an AI agent for data engineering, analytics, ML and agent-building tasks, and it understands your RBAC and schemas. It comes in three forms. In Snowsight, it is built into Workspaces and Admin pages. It writes and explains SQL and Python, answers account questions such as credit consumption, knows which file you have open, and shows a diff view so you can accept changes. CoCo Desktop is a standalone IDE. The CLI is an agentic shell for power users. Unlike the Snowsight UI, it can read and write local repositories such as dbt projects. It can also run bash and git, and it supports AGENTS.md, skills and MCP servers.
Snowflake Copilot inline is a lighter assistant. It needs no setup. In a Workspace you press CMD+I and type a request. Copilot replies inline and shows a diff against your code, which you can accept or reject. Each session is tied to one Workspace file. To work out what data is available to query, it uses the names of databases, schemas, tables and columns and the column data types, so meaningful names help it. This is a developer aid for writing SQL in your editor. Cortex Analyst, by contrast, serves business users through applications and is backed by a semantic view.
Checkpoint 3 of 7· Check yourself
A data engineer wants an AI agent that edits the dbt project in their local Git repository and runs SQL against Snowflake. Which experience supports local file access?
The documentation contrasts the CLI with the Snowsight UI: only the CLI reads and writes local repositories.
“Unlike the Snowsight UI, the CLI can read and write to your local repositories”Source: docs.snowflake.com
4.Interfaces: AI & ML Studio, SQL, Python and REST API
Most Cortex features can be reached through more than one interface, and the exam expects you to match the interface to the situation. The AI & ML Studio in Snowsight (AI & ML in the navigation menu) offers guided, no-code creation; Cortex Search services, for example, can be built there through a wizard. SQL is the interface for batch work and for defining objects. AI Functions are SQL functions, a search service is a CREATE statement, and agents have their own SQL commands. Python gives the same capabilities to notebooks and applications. The REST API is for low-latency, external integration: the Complete, Embed and Agents APIs, the Cortex Analyst API, and agent:run for calling a saved agent.
| Interface | What you do there |
|---|---|
| Snowflake AI & ML Studio (Snowsight) | AI & ML » Cortex Search » Create: a guided wizard for role, warehouse, source data and target lag |
| SQL | CREATE CORTEX SEARCH SERVICE; preview results with SNOWFLAKE.CORTEX.SEARCH_PREVIEW |
| Python API | Query the service from application code with the search method |
| REST API | Query the service from any application over HTTP |
resp = transcript_search_service.search(
query="internet issues",
columns=["transcript_text", "region"],
filter={"@eq": {"region": "North America"} },
limit=1
)
print(resp.to_json())Checkpoint 4 of 7· Check yourself
A team has built a Cortex Agent in Snowsight and now wants to call it from their own customer-facing web application. Which interface do they integrate with?
You can create agents in Snowsight, with SQL or through the REST API, but applications integrate with them through the REST API.
“then integrate it into your application using the REST API.”Source: docs.snowflake.com
Checkpoint 5 of 7· Exam question
A retail analytics team wants business users to type plain-English questions like 'What were total sales by region last quarter?' and receive accurate SQL-generated answers against a governed semantic view. Which Snowflake Gen AI feature should they implement?
Correct answer: C — Cortex Analyst, which converts natural language into SQL over a defined semantic model
- A. This is incorrect: hybrid keyword and vector search is designed for retrieving passages from unstructured text corpora, not for generating SQL against structured sales tables.
- B. This is incorrect: adapting model weights through labeled training examples does not by itself translate a natural-language question into governed SQL against a semantic view.
- C. This is correct: this capability grounds natural-language questions in a defined semantic model and generates the corresponding SQL, which is exactly the chat-with-data pattern described.
- D. This is incorrect: GPU-backed containerized hosting addresses custom model serving infrastructure, not natural-language-to-SQL translation over business data.
5.Bringing your own models: Model Registry and Snowpark Container Services
Cortex gives you Snowflake-hosted models. Snowflake ML covers the opposite case, where you develop and run your own models while the data stays inside Snowflake. The Snowflake Model Registry can log built-in model types directly. For a model whose framework the registry does not support natively, you wrap it in snowflake.ml.model.custom_model.CustomModel. Serialised models trained with external tools or taken from open-source repositories work this way. A ModelContext packages the model files, and you can add code with code_paths. You then log the custom model and deploy it for inference.
For low-latency serving, any model in the registry can be deployed as a managed service with its own HTTP endpoint. Snowflake hosts it as an HTTP server in Snowpark Container Services (SPCS). You do not manage Docker images or Kubernetes clusters, and large models can run on distributed GPU clusters with autoscaling and built-in observability. Before you start you need a model logged in the registry and an understanding of SPCS compute pools and their privileges.
Checkpoint 6 of 7· Match them up
Match each requirement to the Snowflake capability that meets it
Tap a term, then the definition that fits it.
CustomModel brings non-native model types into the registry. SPCS hosts registered models as scalable HTTP services, including on GPU clusters. Fine-tuning adapts a Cortex base LLM rather than importing your own model.
“Serializable models trained using external tools or obtained from open source repositories can be used with CustomModel.”Source: docs.snowflake.com
Checkpoint 7 of 7· Exam question
A support organization has thousands of unstructured PDF manuals and wants a fully managed retrieval layer that combines semantic vector matching with keyword relevance for a RAG-based chatbot. Which feature best fits this requirement?
Correct answer: A — Cortex Search, a managed hybrid search service for unstructured content
- A. This is correct: this service is purpose-built to combine vector-based semantic matching with keyword relevance over indexed unstructured content, which is the retrieval layer a RAG chatbot needs.
- B. This is incorrect: this layer is designed to translate natural-language questions into SQL against structured semantic models, not to retrieve passages from unstructured PDF manuals.
- C. This is incorrect: adapting an LLM's weights through a training job does not provide the indexing and hybrid retrieval mechanics needed to ground answers in a document corpus.
- D. This is incorrect: a model version catalog manages deployed custom models, but it does not perform hybrid semantic and keyword search over unstructured documents.
Exam traps
Each one states something that sounds right. Open it to see what is actually true.
1.Every Cortex Agent tool keeps traffic inside Snowflake's perimeter.Why is that wrong?
Most tools run inside Snowflake, but web search and MCP connectors reach external services over the public internet.
Covered in Cortex Agents: orchestrating Search, Analyst and other tools
2.Only models from frameworks the Model Registry natively supports can be logged and served in Snowflake.Why is that wrong?
Other model types can be wrapped in the CustomModel class, logged to the registry and deployed for inference.
Covered in Bringing your own models: Model Registry and Snowpark Container Services
Sources
Every claim above is drawn from one of these pages, quoted as it was written on the date shown.
- 1.
“An agent reasons over a request, plans the work, calls tools, executes code, and generates a response”
↩︎ Cortex Agents: orchestrating Search, Analyst and other tools“An agent with no tools configured can still hold a conversation”
↩︎ Cortex Agents: orchestrating Search, Analyst and other tools“users can interact with them in Snowflake CoWork and Cortex Code”
↩︎ Snowflake Intelligence and the business-user experience“Tools that reach external services, such as web search and MCP connectors, send requests over the public internet.”
↩︎ Exam trap 1“Reflect and respond: The agent evaluates the results from each tool to decide what to do next”
↩︎ Checkpoint“then integrate it into your application using the REST API.”
↩︎ Checkpoint - 2.
“It lets users interact with their structured and unstructured enterprise data using natural language.”
↩︎ Snowflake Intelligence and the business-user experience“Snowflake CoWork is a ready-to-use agentic application with an intuitive, conversational interface that helps business users discover and act on deep insights.”
↩︎ Checkpoint - 3.
“CoCo is delivered through three experiences: in Snowsight, as a standalone desktop IDE, and as a command line interface (CLI)”
↩︎ Cortex Code in Snowsight, the Cortex Code CLI and Snowflake Copilot inline“Unlike the Snowsight UI, the CLI can read and write to your local repositories”
↩︎ Checkpoint - 4.
“Each session with Snowflake Copilot inline is associated with a particular file in your Workspace.”
↩︎ Cortex Code in Snowsight, the Cortex Code CLI and Snowflake Copilot inline“uses the names of your databases, schemas, tables, and columns and also the data types of your columns”
↩︎ Cortex Code in Snowsight, the Cortex Code CLI and Snowflake Copilot inline - 5.
“These are available for simple inference (Complete API), embedding (Embed API) and agentic applications (Agents API).”
↩︎ Interfaces: AI & ML Studio, SQL, Python and REST API - 6.https://docs.snowflake.com/en/user-guide/snowflake-cortex/cortex-search/cortex-search-overviewOfficial docs
“You can create a Cortex Search Service with a single SQL query or from the Snowflake AI & ML Studio.”
↩︎ Interfaces: AI & ML Studio, SQL, Python and REST API - 7.https://docs.snowflake.com/en/developer-guide/snowflake-ml/inference/real-time-inference-rest-apiOfficial docs
“Snowflake simplifies the deployment pipeline by hosting your model as an HTTP server within Snowpark Container Services (SPCS).”
↩︎ Bringing your own models: Model Registry and Snowpark Container Services“Run large-scale models on distributed GPU clusters for high-performance requirements.”
↩︎ Bringing your own models: Model Registry and Snowpark Container Services - 8.
“Snowflake ML lets you develop and operationalize custom models to solve your unique data challenges, while keeping your data inside Snowflake.”
↩︎ Bringing your own models: Model Registry and Snowpark Container Services
Also cited
- https://docs.snowflake.com/en/developer-guide/snowflake-ml/model-registry/bring-your-own-model-typesOfficial docs
“We also provide a method of logging other model types with snowflake.ml.model.custom_model.CustomModel.”
↩︎ Exam trap 2“Serializable models trained using external tools or obtained from open source repositories can be used with CustomModel.”
↩︎ Checkpoint