Follow these steps to create a function in the agent:

​
Step-by-step guide

​
1. Find “Functions” on the sidebar

  • Go to the agent’s main page and locate “Functions” on the sidebar.

​
2. Click “Add Function”

  • Select “Add Function” on the left side of the page.

​
3. Define the function name

  • At the top of the page, define the function name. Note that the function name can only contain alphanumeric characters and underscores.
Function names are processed by the LLM in the same way any other text, so descriptive accuracy is key. Name functions to explicitly describe what they do and avoid “start” and “stop language”. For example changing start_package_upgrade to get_available_packages prevents the LLM from assuming an unnecessary process is starting.

​
4. Provide a description

  • Use the “Description” field to provide an accurate summary of what your function does. This helps the model understand when to call the function. Be descriptive and concise about its purpose.

​
5. Define LLM parameters

  • In the “LLM Parameters” field, specify the parameters the LLM model will collect and use in the function.
    • Name: Assign a clear and descriptive name to improve the accuracy of the LLM result.
    • Context Description: Provide essential context to help the LLM accurately understand and extract the parameter from the caller.
    • Type: Specify the parameter type. Options include “string,” “number,” “integer,” or “boolean.” Note that “number” supports decimals, while “integer” does not.

​
6. Define the function

  • Use the “Function Definition” field to write the function in Python. You can retrieve “Secrets” in your Python definition—refer to the “How to Retrieve Secret” page for details.
  • The function should return a string, which serves as additional information for the LLM to continue the conversation. Ensure the output is a natural language string that resolves knowledge gaps for the model.

​
7. Save the function

  • Click “Save” to finalize and create your function.

​
Best practices

​
Environment configuration

You can use the conv.env property to define environment-specific functions and activate test features in sandbox or pre-release environments. 
if conv.env == "sandbox":
    <enable_early_feedback_flow>
elif conv.env == "pre-release":
    <activate_staging_tools>
elif conv.env == "live":
    <run_production_features>

​
Handling API response errors

When writing functions that call external APIs (e.g. to fetch user records or transactions), always handle non-200 HTTP responses explicitly.
If response.status_code != 200, make sure to distinguish:
  • A true error (e.g. invalid URL or broken integration)
  • From a valid request with no user data (e.g. user has no transaction history)
Instead of returning a generic failure, you can hand off the call or give a clearer response depending on the cause.
Example: 
response = requests.get(url)

if response.status_code == 404:
    return {
        "handoff": {
            "reason": "no_transaction_data",
            "utterance": "It looks like there's no recent activity on your account. Let me transfer you to someone who can help."
        }
    }
elif response.status_code != 200:
    return {
        "utterance": "Sorry, we couldn't reach your account information right now. Please try again later."
    }