Understand Agent Creation
The Agent Creation feature is integrated into the platform’s workflow configuration settings. The process is organized into several key sections to help you create, configure tools for, and deploy Agents.
Create an Agent
- Within your app, click Intelligence in the left navigation menu.
- Click + Agent in the top right corner.
- Enter a name and description in the respective fields.
- Select an LLM model from the dropdown. If none are available, click Set up AI Service.
- Choose the permissions level.
- Run as publisher: The agent always has the same level of access, regardless of the user interacting with the agent.
- Run as current user: Permissions change with the user interacting with the agent.
- Run as service account: Set up a dedicated account with specific permissions for the agent. This is ideal for integrations or automations that require consistent, controlled access.
- Select the Agent type.
- Click Create Agent.
Test the agent in the chat panel as you update the settings.
- Click the Configure button on the agent profile.
- Add the first message you want the user to see or hear when using the agent.
- Add instructions. This is the prompt the agent will follow. Make sure to include the agent’s overall purpose, core responsibilities, guidance on using specific tools, and any constraints.
Use an LLM to help refine the agent’s prompt for clarity and completeness.
Starting Actions
You can configure Starting Actions—clickable chips that appear below the agent’s greeting before the user’s first message. Each action has a name, icon, and prompt template; optional {{variable}} placeholders open an inline form to collect inputs (Text, Number, or Dropdown) before sending. Actions without variables send immediately. Chips can be reordered by drag-and-drop and toggled on or off. For full setup and variable options, see Agent Starting Actions.
When the agent runs a tool, users see a live status message in the chat showing that a tool is executing. Admins can set a custom display message per tool (e.g., “Searching knowledge base…”) so users understand what is happening. Multiple back-to-back tool calls appear as separate messages. This behavior is supported in both the agent builder preview and deployed app agents.
Add and configure tools for Agents to perform actions like creating or searching records, running automations, delegating to sub-agents, or connecting to external MCP services.
- In the Agent profile page, click the Configure button.
- Click + Tool in the Tools section.
- Select the tool type.
- Complete the fields, including:
- Name and description
- Permissions
- Object access
- Number of results allowed
- Conditions
- Output fields
- Save the tool.
- Test the tool and make changes as needed.
Allows the Agent to create new records in the system.
- Tool Start Message: Customize what the user sees or hears when the tool runs.
- Run As: Select execution permissions — run as publisher (default), run as current user, or run as service account.
- App/Element/Task: Choose the specific App, Element, or Task that the Agent can create.
- Fields: Required fields are shown by default. Add optional fields as needed, and provide a description for each field to guide the Agent on how to obtain the necessary data.
- Return Fields: Choose which fields from the created record should be sent back to the Agent.
Enables the Agent to search for existing records based on a query.
- Tool Start Message: Customize what the user sees or hears when the tool runs.
- Run As: Select execution permissions — run as publisher (default), run as current user, or run as service account.
- App/Element/Task: Choose the App, Element, or Task to search.
- Query Description: Provide guidance on what the Agent should search for.
- Return Fields: Choose which fields from the search results should be returned to the Agent.
- Input value references (Beta): Define input parameters with descriptions that guide the agent. Reference these inputs in tool filters; the agent passes values into the input parameters at runtime, so filters can be driven by live context instead of static configuration.
The search returns results similar to a list view search. If more than 100 records are found, the Agent receives a failure message indicating the need for a more refined query.
Allows the Agent to update existing records.
- Tool Start Message: Customize what the user sees or hears when the tool runs.
- Run As: Select execution permissions — run as publisher (default), run as current user, or run as service account.
- Record Identification: Explain how the Agent should identify the record to update (e.g., by record ID).
- App/Element/Task: Choose the App, Element, or Task to update.
- Fields to Update: Select the fields the Agent must update, and provide descriptions for each field to guide the Agent.
The Agent is notified whether the update was successful or not, along with reasons for failure, allowing it to attempt corrections.
Provides advanced search capabilities using AI, allowing the Agent to understand and retrieve related data based on natural language.
- Tool Start Message: Customize what the user sees or hears when the tool runs.
- Run As: Select execution permissions — run as publisher (default), run as current user, or run as service account.
- App/Element/Task: Choose the App, Element, or Task to search.
- Result Limit: Specify the maximum number of results to return (1–100, default is 10).
- AI Search Configuration: Select the AI Search configuration for the object. Choose attribute fields to search on and set up any necessary filters. Provide descriptions for each field to guide the Agent.
- Input value references (Beta): Define input parameters with descriptions that guide the agent. Reference these inputs in tool filters; the agent passes values into the input parameters at runtime, so filters can be driven by live context instead of static configuration. See AI Search for configuring input parameters when using AI Search as an agent tool.
- Return Fields: Choose which fields from the search results should be sent back to the Agent.
Allows the Agent to execute existing automations that have an On-Demand trigger, enabling complex workflows to be triggered conversationally.
- Tool Name: No spaces or special characters allowed.
- Tool Description: Describe when and how the Agent should use this tool.
- Tool Start Message: Customize what the user sees or hears when the tool runs.
- Run As: Select execution permissions — run as publisher (default), run as current user, or run as service account.
- Automation: Choose an existing automation with an On-Demand trigger.
- Input Values: Specify the input parameters the Agent will pass to the automation, including field names and descriptions.
- Output Values: Specify which output values the Agent will receive from the automation.
The automation executes with the configured permissions. Input values are collected from the conversation context based on field descriptions. Output values are returned to the Agent for use in the conversation.
Allows the Agent to delegate tasks to another agent (sub-agent) for specialized processing. Each execution creates a new conversation between the main agent and the sub-agent.
- Tool Name: No spaces or special characters allowed.
- Tool Description: Describe when and how the Agent should use this tool.
- Tool Start Message: Customize what the user sees or hears when the tool runs.
- Run As: Select execution permissions — run as publisher (default), run as current user, or run as service account.
- App: Choose the App containing the target agent.
- Target Agent: Choose the sub-agent to delegate tasks to.
- Worker Task Prompt: Provide instructions for how this agent should delegate tasks to the worker agent.
Each tool execution starts a new conversation with the sub-agent. The main agent can provide context and instructions, and results from the sub-agent are returned for continued processing.
Allows the Agent to connect to external MCP servers and use custom tools, extending the Agent’s capabilities beyond the platform’s built-in tools.
Configuration follows a 3-step process:
Step 1: Connect to MCP Server
- MCP Server URL: Enter the base URL of the MCP server (e.g.,
https://api.example.com/mcp).
- Authentication: Select the authentication method:
- Bearer Token: Enter an access token.
- Basic Auth: Enter username and password.
- API Key: Enter the API key.
- None: No authentication required.
- Click Connect to discover available tools.
Step 2: Select Tool
- Browse and select from the tools available on the MCP server.
Step 3: Configure Tool
- Tool Name: No spaces or special characters allowed.
- Tool Description: Describe when and how the Agent should use this tool.
- Tool Start Message: Customize what the user sees or hears when the tool runs.
- Run As: Select the execution permissions.
- Click Save Tool to complete the configuration.
The Agent communicates with the external MCP server to execute the tool. Authentication is handled automatically based on the configured method. Results are returned to the Agent for use in the conversation.
Add Your Agent to Your App Workflow
Once your agent is created and configured, you can edit and test by returning to Intelligence in your app.
The agent can now be added into Automations. You can also use the agent whenever you see the Agent ✨ button in Elementum, through Microsoft Teams, Slack, and through phone integrations.
