Overview

Prerequisites

Part 1: Infrastructure Setup

Step 1: Generate Personal Access Token (PAT)

Step 2: Configure and Run Setup Script

-- Step 1: Network Rule (egress to Snowflake API)
CREATE OR REPLACE NETWORK RULE cortex_agent_egress_rule ...

-- Step 2: Store PAT Token as Secret
CREATE OR REPLACE SECRET cortex_agent_token_secret ...

-- Step 3: External Access Integration
CREATE OR REPLACE EXTERNAL ACCESS INTEGRATION cortex_agent_external_access ...

-- Step 4: Grant Permissions
GRANT READ ON SECRET ... TO ROLE <YOUR_ROLE>;
GRANT USAGE ON INTEGRATION ... TO ROLE <YOUR_ROLE>;

-- Step 5: Create Agent Caller UDF (template)
CREATE OR REPLACE FUNCTION call_cortex_agent(user_query STRING) ...

Step 3: Create a UDF for each sub-agent

-- Agent 1
CREATE OR REPLACE FUNCTION ASK_AGENT_1(user_query VARCHAR)
RETURNS STRING
LANGUAGE PYTHON
RUNTIME_VERSION = '3.12'
PACKAGES = ('requests', 'snowflake-snowpark-python')
EXTERNAL_ACCESS_INTEGRATIONS = (cortex_agent_external_access)
SECRETS = ('agent_token' = cortex_agent_token_secret)
HANDLER = 'run_agent'
AS
$$
...
url = "https://<YOUR_ACCOUNT>.snowflakecomputing.com/api/v2/databases/<DB>/schemas/<SCHEMA>/agents/AGENT_1:run"
...
$$;

-- Agent 2
CREATE OR REPLACE FUNCTION ASK_AGENT_2(user_query VARCHAR)
RETURNS STRING
...
url = "https://<YOUR_ACCOUNT>.snowflakecomputing.com/api/v2/databases/<DB>/schemas/<SCHEMA>/agents/AGENT_2:run"
...
$$;

Part 2: Master agent setup (UI)

Step 4: Create master agent in Snowflake Intelligence

Step 5: Configure agent description

A multi-domain routing agent that serves as a central interface, automatically delegating user questions to specialized sub-agents based on query intent.

Step 6: Add custom tools

Tool Name:       AGENT_1
Function name:   ASK_AGENT_1(VARCHAR)
Warehouse:       SNOWFLAKE_INTELLIGENCE_WH
Description:     Provides information on [Agent 1's domain]

Step 7: Configure orchestration prompt

You are an intelligent router. Analyze the user's inquiry to determine the correct domain:

Agent 1: Use this for questions regarding [domain 1 topics].

Agent 2: Use this for questions regarding [domain 2 topics].

Agent N: Use this for questions regarding [domain N topics].

Rule: Do not attempt to answer the question yourself. You must strictly invoke the appropriate tool to get the answer.

You are an intelligent router. Analyze the user's inquiry to determine the correct domain:

Agent 1 (AGENT_1): Use this for questions regarding sports, matches, player statistics, teams, or game history.

Agent 2 (AGENT_2): Use this for questions regarding business performance, revenue, orders, customer lists, or sales data.

Rule: Do not attempt to answer the question yourself. You must strictly invoke the appropriate tool to get the answer.

Step 8: Configure Access (Access Tab)

Usage

Via Master Agent (Recommended)

Direct Sub-Agent Calls (Testing/Debugging)

-- Test Agent 1 directly
SELECT ASK_AGENT_1('What is the status of X?');

-- Test Agent 2 directly  
SELECT ASK_AGENT_2('Show me the report for Y');

Examples:

Architecture components

API endpoint format

https://<ACCOUNT>.snowflakecomputing.com/api/v2/databases/<DATABASE>/schemas/<SCHEMA>/agents/<AGENT_NAME>:run

Troubleshooting

Security considerations

Resources

SQL Code

*\-- CORTEX AGENT MULTI-AGENT SETUP (GENERIC TEMPLATE)*  
*\-- \============================================================================*  
*\-- This script sets up the infrastructure to call Cortex Agents programmatically*  
*\-- via UDFs, enabling multi-agent orchestration patterns.*  
*\--*  
*\-- PLACEHOLDERS TO REPLACE:*  
*\--   \<YOUR\_ACCOUNT\>       \- Your Snowflake account identifier (e.g., ABC12345)*  
*\--   \<YOUR\_DATABASE\>      \- Database where agents are defined*  
*\--   \<YOUR\_SCHEMA\>        \- Schema where agents are defined*   
*\--   \<YOUR\_AGENT\_NAME\>    \- Name of your Cortex Agent*  
*\--   \<YOUR\_ROLE\>          \- Role that will execute the agent UDFs*  
*\--   \<YOUR\_PAT\_TOKEN\>     \- Personal Access Token (generated from Snowflake UI)*  
*\-- \============================================================================*

USE ROLE ACCOUNTADMIN;

*\-- \============================================================================*  
*\-- STEP 1: CREATE NETWORK RULE FOR EGRESS TO SNOWFLAKE API*  
*\-- \============================================================================*  
*\-- Allows the UDF to make outbound calls to your Snowflake account's REST API.*  
*\-- Required for the agent to call itself via HTTP.*

CREATE OR REPLACE NETWORK RULE cortex\_agent\_egress\_rule  
 MODE \= EGRESS  
 TYPE \= HOST\_PORT  
 VALUE\_LIST \= ('\<YOUR\_ACCOUNT\>.snowflakecomputing.com');

*\-- \============================================================================*  
*\-- STEP 2: STORE PAT TOKEN AS A SECRET*  
*\-- \============================================================================*  
*\-- Securely stores the Personal Access Token.*  
*\-- Generate PAT from Snowflake UI: User Menu → My Profile → Authentication → Generate Token*  
*\-- NOTE: PAT tokens have an expiration date. Regenerate and update secret as needed.*

CREATE OR REPLACE SECRET cortex\_agent\_token\_secret  
 TYPE \= GENERIC\_STRING  
 SECRET\_STRING \= '\<YOUR\_PAT\_TOKEN\>';

*\-- \============================================================================*  
*\-- STEP 3: CREATE EXTERNAL ACCESS INTEGRATION*  
*\-- \============================================================================*  
*\-- Bridges the network rule and secrets, allowing UDFs to make authenticated*  
*\-- external calls to the Snowflake API.*

CREATE OR REPLACE EXTERNAL ACCESS INTEGRATION cortex\_agent\_external\_access  
 ALLOWED\_NETWORK\_RULES \= (cortex\_agent\_egress\_rule)  
 ALLOWED\_AUTHENTICATION\_SECRETS \= ALL  
 ENABLED \= TRUE;

*\-- \============================================================================*  
*\-- STEP 4: GRANT PERMISSIONS TO EXECUTION ROLE*  
*\-- \============================================================================*  
*\-- Grant the necessary permissions to the role that will call the agent UDFs.*

GRANT READ ON SECRET cortex\_agent\_token\_secret TO ROLE \<YOUR\_ROLE\>;  
GRANT USAGE ON INTEGRATION cortex\_agent\_external\_access TO ROLE \<YOUR\_ROLE\>;

*\-- \============================================================================*  
*\-- STEP 5: CREATE GENERIC AGENT CALLER UDF*  
*\-- \============================================================================*  
*\-- Template UDF that calls a Cortex Agent via REST API.*  
*\-- Duplicate and customize for each agent you want to call.*

CREATE OR REPLACE FUNCTION call\_cortex\_agent(user\_query STRING)  
RETURNS STRING  
LANGUAGE PYTHON  
RUNTIME\_VERSION \= '3.12'  
PACKAGES \= ('requests', 'snowflake-snowpark-python')  
EXTERNAL\_ACCESS\_INTEGRATIONS \= (cortex\_agent\_external\_access)  
SECRETS \= ('agent\_token' \= cortex\_agent\_token\_secret)  
HANDLER \= 'run\_agent'  
AS  
$$  
import \_snowflake  
import requests  
import json

def run\_agent(user\_query):  
   """  
   Calls a Cortex Agent via REST API and returns the text response.  
   Handles Server-Sent Events (SSE) streaming format.  
   """  
   \# Retrieve the stored PAT token  
   try:  
       token \= \_snowflake.get\_generic\_secret\_string('agent\_token')  
   except Exception as e:  
       return f"Error: Could not read secret. Verify grants. Details: {str(e)}"

   \# Construct the agent API endpoint  
   \# Format: https://\<account\>.snowflakecomputing.com/api/v2/databases/\<db\>/schemas/\<schema\>/agents/\<agent\>:run  
   url \= "https://\<YOUR\_ACCOUNT\>.snowflakecomputing.com/api/v2/databases/\<YOUR\_DATABASE\>/schemas/\<YOUR\_SCHEMA\>/agents/\<YOUR\_AGENT\_NAME\>:run"  
    
   headers \= {  
       "Authorization": f"Bearer {token}",  
       "Content-Type": "application/json",  
       "Accept": "text/event-stream"  
   }  
    
   payload \= {  
       "messages": \[  
           {  
               "role": "user",  
               "content": \[{"type": "text", "text": user\_query}\]  
           }  
       \]  
   }  
    
   try:  
       \# Make streaming request to agent API  
       response \= requests.post(url, headers\=headers, json=payload, stream\=True)  
        
       if response.status\_code \!= 200:  
           return f"API Error {response.status\_code}: {response.text}"  
        
       \# Parse Server-Sent Events (SSE) stream  
       final\_answer \= \[\]  
       current\_event \= None  
        
       for line in response.iter\_lines():  
           if not line:  
               continue  
            
           decoded\_line \= line.decode('utf-8')  
            
           \# Track event type  
           if decoded\_line.startswith('event: '):  
               current\_event \= decoded\_line\[7:\].strip()  
            
           \# Extract data payload  
           if decoded\_line.startswith('data: '):  
               data\_str \= decoded\_line\[6:\]  
               if data\_str \== '\[DONE\]':  
                   break  
                
               try:  
                   data \= json.loads(data\_str)  
                   \# Collect text delta events (the actual response content)  
                   if current\_event \== 'response.text.delta' and 'text' in data:  
                       final\_answer.append(data\['text'\])  
               except json.JSONDecodeError:  
                   continue  
        
       return "".join(final\_answer) if final\_answer else "Agent returned no text content."

   except Exception as e:  
       return f"Connection error: {str(e)}"  
$$;

*\-- \============================================================================*  
*\-- EXAMPLE: CREATE ADDITIONAL AGENT CALLERS*  
*\-- \============================================================================*  
*\-- Copy and modify the UDF above for each agent. Only change the URL endpoint.*  
*\--*  
*\-- Example for a second agent:*  
*\--*  
*\-- CREATE OR REPLACE FUNCTION call\_second\_agent(user\_query STRING)*  
*\-- RETURNS STRING*  
*\-- ... (same as above, but change the url variable to point to different agent)*

*\-- \============================================================================*  
*\-- USAGE EXAMPLES*  
*\-- \============================================================================*  
*\-- SELECT call\_cortex\_agent('What is the current status?');*

*\-- \============================================================================*  
*\-- OPTIONAL: NETWORK POLICY MODIFICATION (IF REQUIRED)*  
*\-- \============================================================================*  
*\-- If your account has network policies blocking agent traffic, you may need*  
*\-- to add Cortex Agent IP ranges. Check with your admin first.*  
*\--*  
*\-- \-- View current network policies*  
*\-- SHOW PARAMETERS LIKE 'NETWORK\_POLICY' IN ACCOUNT;*  
*\-- SHOW PARAMETERS LIKE 'NETWORK\_POLICY' IN USER \<YOUR\_USERNAME\>;*  
*\--*  
*\-- \-- If modification is needed, add required IPs to allowed list*  
*\-- ALTER NETWORK POLICY \<YOUR\_POLICY\_NAME\> SET ALLOWED\_IP\_LIST \= (*  
*\--     ... existing IPs ...,*  
*\--     '\<AGENT\_IP\_RANGE\>'  \-- Add agent IPs here*  
*\-- );*