Create and manage AI Agent

This topic describes how to create and manage an AI Agent in the Control Room. You can perform various actions such as running, moving, checking in, viewing, editing, copying, and deleting AI Agent, with certain permissions required for editing and deletion.

To create an AI Agent in the Control Room:

Prerequisites

Ensure you have the necessary 역할 및 권한 configured.

Procedure

  1. Log in to the Control Room and navigate to Automation > Create new or ‘+’ icon and choose AI Agent.
  2. Provide a name and description and click Create & edit to display a template outline. For example, this topic uses a Claim Details AI Agent.
  3. Define the agent's prompt: The Generate prompt for your Agent screen (cloud-only feature) appears, prompting you to enter the prompt.

    Define the agent’s mission by providing a clear, detailed prompt that describes the agent’s purpose in natural language. You will do this by providing a clear and detailed initial screen prompt that describes the agent's purpose in natural language. This prompt is the foundation upon which the entire agent's logic is built.

    For example, to create the Claim Details agent shown in the example, a user might have started with a concise but informative message like this:

    Your job is to retrieve claim details:
*  Retrieve policy and police report details.
*  If any information is missing, ask a human for assistance.

    This initial input provides the core requirements, which the AI Agent Studio can then use to generate a more detailed, structured prompt. The AI Agent Studio is designed to be flexible; you do not need to provide every single detail upfront. However, the more context you provide, the better the AI Agent Studio can generate a starting template for you to refine.

    The system will then expand on this initial input to create a comprehensive prompt (example shown below).

    Note: The generated role, goal, and action plan text varies slightly based on the model. Your output will not match this word-for-word, but the structure and intent will be similar.

    Example output:

    Role: You are an insurance claims processor specializing in stolen property claims, with expertise in retrieving and verifying claim, policy, and police report details using automated tools and human assistance when necessary.

    Goal:Your goal is to retrieve and verify all required details for a stolen property claim using the provided claim ID, ensuring no output value is null, and consulting a human if necessary to complete the task.

    AI Agent- Claim Details
  4. Define the ending states: In the Ending state definition section of the Prompt tab, specify the status messages that the agent reports when execution ends. At least one state is required.

    The three available ending states are:

    • Completed: The agent successfully achieved its goal. Use this state to confirm successful task completion. Example: This state must be set if you were able to achieve your goal.
    • Failed: The agent was unable to complete its goal. Use this state to communicate failure to the calling process or user. Example: This state must be set if you were unable to achieve your goal.
    • Cancelled: The agent stopped because it received a cancellation instruction from the user or a tool. Select Enabled cancelled status to activate this state. Example: This state must be set if you received instructions from the user or a tool to cancel your efforts to achieve your goal.
      Note: The Cancelled state is optional. Select the Enabled cancelled status check box to activate it. If you do not enable this state, the agent will return either Completed or Failed at the end of execution.

    The text you enter in each state field is the message passed back to the calling automation or Automation Co-Pilot interface when the agent finishes. Provide clear, descriptive messages so that downstream processes can handle each outcome appropriately.

  5. Choose a Model Connection:Navigate to Model tab and click Choose to choose from the available list of Model connections you have access to. For more information on Model Connection, see 모델 연결
    These Model connections are created by the Automation Admin and assigned to your user with a custom role. Only Model connections created from Automation Anywhere provided models or from bring-your-own-licenses (BYOL) are allowed. Custom model definitions are not currently supported with AI Agent.
    Provider Supported Models (Standard type)
    OpenAI & Azure OpenAI GPT-4o, GPT-4.1
    Amazon Bedrock Anthropic Claude 4.5 (Sonnet, Haiku, Opus)
    Google Vertex AI Anthropic Claude 4.5 (Sonnet, Haiku, Opus)
    Automation Anywhere GPT-4o, Claude Sonnet 4.5
  6. Define agent variables: Navigate to Input & Output tab. Here, you need to tell the agent what information to look for as input variables and what to provide as output variables. These variables will be used by the LLM to understand the context and communicate with the tools. The key is providing rich and descriptive language. The more detailed your descriptions, the better the LLM will be at understanding its purpose and fulfilling its goal.

    • Input variable:

      • Name: claim_id

      • Description: The claim ID for which to retrieve details

    • Output variable:

      • Name: claim_details

      • Description: Final details of the claim, including policy and police report information.

  7. Define the End-user Display: Configure how AI Agent execution is presented to the user. The goal is to provide a clear and meaningful title for the task that the user sees in the Co-pilot interface. Navigate to the End-user display tab. In the Agent request title, provide a title for the task. This title is a mandatory field and is displayed when the AI Agent runs in the Co-pilot. For example, for the Claim Details agent, the title could be Claim Details Request. This helps the user understand what the agent is doing at a glance.
  8. Add and configure tools:

    Under Tools tab, add the required automations for your agent to perform its task. In this example, you will be adding the Claim Lookup, Police Report, Data Tracker, and Policy Process automations as tools.

    Tool Name Purpose Input Variables Output Variables
    claim lookup Provides the claim details. INPUT_id: claim id claim_id: the claim id, policy_id: the policy id, stolen_property_item: the item that was stolen, claimed_replacement_cost: the replacement cost for the claim items, location_of_incident: location of the incident, police_report_id: the police report id, date_of_incident: the date of the incident, claim_status: the status of the claim, claimant_name: the claimant name, receipt: the receipt for the property in the claim
    police report Gets the police report details. INPUT_id: police report id police_report_id: the police report id, report_date: Date of report, officer_name: name of officer, officer_badge_number: badge number, incident_type: type of incident, description: description of input, incident_location: incident location, suspect_name: suspect name, arrest_made: Was an arrest made, reporting_station: reporting station
    data tracker Always calls this tool to track data you have or have not received so far data_received_or_missing_so_far: The details on the data you have received or are missing (or have null values) formatted in json data_received_or_missing_so_far_response: The data you have received or are missing so far in JSON format
    policy process This gets the details of a policy INPUT_id: The policy id policy_id: the policy id, policy_number: the policy number, policy_type: the policy type, issue_date: The date the policy was issued, expiration_date: the expiration date, status: status of the policy, premium_amount: the premium amount, payment_frequency: the frequency of payments, insured_name: the name of the insured, insured_address: the address of the insured, insured_phone: the phone number of the insured, insured_email: the email of the insured, coverage_type: the coverage type, policyholder_id: The policy holder id, coverage_amount: the coverage amount, agent_id: the id of the insurance agent, deductible_amount: the deductible amount, was_successful: True if the lookup found any data, False otherwise.

    For more information on configuring tools, see AI Agent의 도구 설정 구성.

  9. Define the Action Plan: The Action Plan section, is generated by the system based on your prompt and the tools you have added. It outlines the logical flow:
    • Retrieve claim details: Call the Claim Lookup tool using the claim_id.
    • Retrieve police report: Call the Police Report tool.
    • Track the data: call the Data Tracker tool.
    • Get policy details: call the Policy Process.
    • If claim details are incomplete, use the Human in the Loop to ask for assistance.
  10. Test and refine: Once you have configured everything, click Run to test your agent to ensure it behaves as expected. You can simulate a conversation with the agent, providing a claim ID and seeing if it correctly retrieves the claim details, police report, and policy information. Refine the variable descriptions and tool prompts as needed to improve the agent's performance and accuracy.
    Note:

    Testing and running AI Agents in your private workspace does not consume Automation AI Credits. For more information on credit consumption when moving to production, see AI Agents 라이선스 모델.

    .

Next steps

Once you have created and configured your AI Agent, embed it into your business processes and bring intelligence into end-to-end automation. For more information on adding an AI Agent to a process, see 프로세스 자동화에 AI Agent 추가. The following diagram illustrates the typical workflow, from agent creation to execution in a public environment.

AI Agent Work flow

The process of moving an AI Agent from a development environment to a public-facing one involves several key steps and user roles.

Create and test
A professional developer or automation admin creates and tests the AI Agent.
Configure in a process
The Professional developer then configures the AI Agent within a larger process.
Check in
The process is checked in to the repository, making it available for use.
Run in public
The process can then be run in a public workspace, allowing it to be called through the Automation Co-Pilot.
Note:

A business user with the required Automation Co-Pilot Business User license can execute the process and interact with the AI Agent. The diagram also shows the specific permissions required for each user role to perform these tasks. For more information about the roles and permissions, refer 역할 및 권한.