An action marks a specific task that the agent must undertake, enabling the agent to gather necessary data or information from the input and activate a tool. The tool is responsible for executing tasks such as summarizing information, generating reports, creating content, or automating workflows, thereby assisting agents in achieving their designated goals.
Actions enable Agents to perform complex operations in real time as part of their goal-driven workflow.
The prompt fields are a critical step that will determine how to process and execute tasks.
New Action
Fill the presented fields:
Name: It should be clear and descriptive of the task the action will perform. Use lowercase letters and separate words with underscore (e.g., search_flight, destination_recommender).
Description: This field provides detailed information about what the action does, its purpose, and when it should be triggered, which makes it essential for its correct functioning. For the action search_flight, for instance, the description is “When users request to book a flight, buy a ticket, reserve a trip, or similar actions, execute them accordingly.”
Parameters
Once you've defined what it should do, now specify how. Breakdown which data should be collected by using the Parameters block. Start by choosing between two approaches: Basic and Advanced.
In advanced mode, you can use a JSON containing all the parameters required for the success of this action for more technical and precise action definition. Whereas in basic mode, you can add the parameters individually for easier configuration.
Fill the presented fields:
Parameter Name
Type: Defines the data format (string, boolean, number) that the parameter accepts
Parameter Description: Explains the purpose and expected values for this specific parameter
Variable (optional): Assigns a reference name that can be used elsewhere in the agent's workflow. The value can be accessed through $hiddenContext.variable, such as: $hiddenContext.numberOfPassengers (see table).
Rules (optional): Specifies constraints, validations, or conditional behaviors that govern how the parameter works. They serve as tactical directives.
For the action search_flight abovementioned, there are five parameters: date, origin, destination, type of flight, number of passengers. Here's a more detailed breakdown for two of them:
Name
numberOfPassengers
dates
Type
String
String
Description
The number of passengers traveling, including both adults and children. The passengers must be adults.
The dates of travel, including departure and optionally return, provided in a format like YYYY-MM-DD or DD/MM/YY. The date provided is a valid date, i.e. it must be a future date, The departure and destination dates must correspond to a valid period. The current year is 2025.
Variable
numberpassengers
--
Rules
- If you are informed that they are a children or student, kindly recommend that they book the flight online: website.com/buy-tickets
- The number of passengers must be 9 people max. If there are more than 9 people, kindly recommend that they book the flight online: website.com
If the user gives you a round-trip date, assume it is a 'round-trip flight'
Note on Data Persistence: When an Action successfully captures specific data, this information is automatically preserved in the conversation context. This persistence creates an important consideration for multi-agent workflows:
If you want a subsequent agent to explicitly verify or reconfirm this information with the user, you must clearly state this requirement in that agent's instruction field. Without such explicit instructions, the following agent will assume this information is already validated and will suppress redundant questions or confirmations.
These fields collectively form the framework that enables an agent to properly understand, process, and respond to user inputs while executing the specific tasks needed to achieve its goals.
Tools
If you need to execute an API call, add a Service cell right after after the desired action in the workflow. It can be a Webhook or a Rest Connector cell, mechanisms that transforms the Action's intent into concrete API interactions, allowing the agent to retrieve or manipulate data needed to fulfill its purpose.
In the example, a Rest cell was chosen to perform an API call. At a visual level, you’ll see the Agent cell and the branches for each Action added and for error handling. In this case, itinerary_option and destination_recommender, and the button error.
SCAI Workspace
Action templates
Examples of actions to speed up your agent creation process.
See list of templates
Customer Satisfaction Survey
Name: collect_satisfaction_feedback
Description: Request an evaluation of the experience and a rating from 1 to 5, and then sends this information to the customer service. Asks if there is any optional comments they may want to share.
Parameters:
rating (number): User satisfaction score. A number from 1 to 5, where 1 is very dissatisfied and 5 is very satisfied.
comments (string): Text input with feedback.
recommend_to_others (boolean): Would recommend to others
Customer Lookup or Identification
Name: identify_customer
Description: Request key information to identify the customer, such as name, ID, or contact details. Use the inputs to search for the customer in the system.
Parameters:
customer_id (string): Unique identifier, if available.
email (string): Email address.
phone (string): Phone number.
HMTL Formatting
Name: format_html_response
Description: This action should be triggered whenever the agent is going to deliver a response to the user, with the aim of formatting the HTML content. It organizes the text in a structured and visually clear way, applying HTML styles and markups according to the context of the response, ensuring a richer and more interactive presentation of the information.
Parameters:
hyml_response (string): Always format all answers using HTML. Use the following guidelines to present the content:
1. List items or options using HTML tags `<ul>` and `<li>` to create an organized list.
2. Highlight important names or key terms using the `<strong>` tag for bold.
3. Use paragraphs `<p>` to separate blocks of text, ensuring visual organization.
4. All formatting must be semantically correct in HTML.
5. Responsive resolution (mobile version)
The output must be well structured to be used directly on any HTML page.
Transfer to Human Agent
Name: handover_to_human
Description: Initiate the handover process to a live agent. Collect necessary context and inform the user about the transfer.
Parameters:
reason (string): Reason for the handover.
urgency (string): Optional level of urgency (e.g., low, medium, high).
Offer Search
Name: search_offers
Description: Search for available offers or promotions based on the user’s preferences or profile.
Parameters:
category (string): Type of offer or service area (e.g., internet, mobile, insurance).
location (string): Optional filter by user’s region.
customer_segment (string): Customer type or profile (e.g., new customer, existing, premium).
Schedule technical visit
Name: schedule_technical_visit
Description: Collect necessary data from the user to schedule a technical visit, such as date preferences and address. All visits have a 3-hour window: 9 to 12, from 12 to 15, from 15 to 18 and from 18 to 20. If someone asks to schedule at any moment within any of those windows, inform the schedule will be within this range.
Parameters:
timetable (string): Refers to the start time of one of these ranges: 9-12, 13-15, 16-18. Entries within these ranges must be the initial value.
preferred_date (string): User’s preferred date for the visit. The day must be informed.
address (string): Location for the visit.
issue_type (string): Brief description of the issue to be resolved.
Collect data for API calls (transactional services)
Name: collect_api_data
Description: Gather structured data from the user required to trigger a transactional API call, such as service request, purchase, or account update.
Parameters:
operation_type (string): Type of operation (e.g., payment, subscription update).
required_fields (object): Key-value pairs with the necessary inputs (e.g., {"account_id": "12345", "amount": "50.00
Channel customization - Web
Name: customize_web_response
Description: Format and adjust the response for a web-based chat interface. Prioritize clarity, readability, and proper formatting using HTML or line breaks when needed. Avoid emojis or informal tone.
Parameters:
raw_response (string): Original response to be formatted.
include_links (boolean): Whether to include clickable links in the message.
format_type (string): Optional format style (e.g., paragraph, bullet list).
Channel customization - Facebook/Meta
Name: customize_facebook_response
Description: Adapt the response for Facebook Messenger or Meta platforms. Keep the tone conversational and friendly. Emojis are allowed. Keep messages short and engaging, using quick replies or buttons when appropriate.
Parameters:
raw_response (string): Original response to be adapted.
use_emojis (boolean): Whether emojis should be added.
button_labels (array of strings): Optional list of quick reply buttons to include.
Channel customization - WhatsApp
Name: customize_whatsapp_response
Description: Format the response for WhatsApp, using a concise and natural tone. Use markdown where supported (bold, italics), and avoid long paragraphs. Keep messages compact and mobile-friendly.
Parameters:
raw_response (string): Original response to be customized.
use_markdown (boolean): Whether to apply WhatsApp-compatible markdown formatting.
split_message (boolean): Whether to split long responses into multiple shorter messages.
