Add a task agent step

Add a task agent step to your workflow to perform IT Operations automation tasks such as log analysis, incident triage, configuration drift analysis, or context-aware decision-making.

Before you begin

  • At least one large language model (LLM) provider integration must be configured and enabled. Navigate to Configuration > Integrations to verify.
  • The LLM provider integration must have discovered at least one model. If no models appear, refresh the integration.
  • An LLM provider credential must exist for the integration you plan to use. If no LLM provider credential exists, you can create one during step configuration.

Procedure

  1. Navigate to the Workflows page and select the workflow you want to modify.
  2. Click Add step > Task Agent.

    The task agent details panel opens with the Parameters pane displayed.

  3. Optional: Change the default step name to a meaningful name that reflects its purpose.
  4. Select a model from the Model dropdown list.

    Models are grouped by LLM provider integration. The list displays the integration name as the group header and lists available models under each group. After you select a model, the panel shows the integration name and model name.

  5. Configure the LLM provider credential.

    After you select a model, a credential status section is displayed below the Model dropdown:

    • If no credential is configured, a warning displays with a Set up connection link. Click it to open a credential selector, then select an existing LLM provider credential or click Create new credential to create one inline.
    • If a credential is already configured, a green checkmark displays with the credential name. Click Change to select a different credential.
  6. In the Prompt field, enter natural language instructions that describe what the agent should do.

    Be specific about:

    • What the agent should analyze or produce
    • What format the response should use
    • Any constraints or boundaries for the response
  7. Optional: To inject data from an upstream step to the Prompt field in the center pane, drag or copy and paste the input parameters if Schema is selected, or copy and paste the parameters if JSON is selected.

    Example prompt:

    Use these logs to analyze the server issue and determine which job template should be run to remediate the problem. Provide a brief analysis summary and the job template name.

  8. Use the Tools dropdown list or use the default to control which tools the agent accesses.

    Tools are discovered from enabled Model Context Protocol (MCP) server integrations. If a tool you select is not available at runtime, the engine logs a warning and proceeds with the remaining valid tools.

    • All tools: The agent can call any tool discovered from enabled MCP server integrations. Use when the agent needs to query multiple external systems.
    • No tools: (Default) The agent operates in text-only mode and cannot make any tool calls. Use when the agent's task is purely analytical, such as analyzing logs or classifying alerts.
    • Select specific tools: The agent can call only the tools you select from the available list. Tools are grouped by their MCP server integration. Use this option to constrain the agent to specific capabilities. For example, allow access to a GitHub integration but not an AWS integration.
  9. If you selected specific tools or all tools, configure the Connections section that is displayed below the Tools field.

    This lists each MCP server integration that provides the selected tools and shows whether an execution credential is configured for that integration.

    • If a connection shows a warning icon, click Set up connection to select an execution credential for that integration. The agent uses this credential to authenticate with the MCP server at runtime.
    • If a connection shows a green checkmark, the credential is already configured. Click Change to select a different credential.

    Each integration must have a connected credential before the agent can call its tools. If a credential is missing, disabled, or misconfigured, the workflow fails immediately when you run it.

  10. Optional: In the Response schema field, define a JSON Schema to enforce a structured output.

    Click the code editor to enter your schema inline, or click the expand icon to open a larger editing window. The schema must be a valid JSON object.

    When you define a response schema, the agent returns its response as a structured JSON object that matches your schema. Downstream steps can then reference individual fields from the response using expressions.

    Example response schema:

    {
"type": "object",
"properties": {
    "analysis_summary": {
    "type": "string",
    "description": "Brief summary of the server issue"
    },
    "job_template_name": {
    "type": "string",
    "description": "Name of the job template to run"
    },
    "severity": {
    "type": "string",
    "enum": ["low", "medium", "high", "critical"]
    }
},
"required": ["analysis_summary", "job_template_name"]
}
  11. Optional: In the Context file upload field, click Upload or drag and drop up to 10 files to give the agent additional information beyond its training data.

    Use one of the following file types: PDF, DOC, DOCX, TXT, MD.

    When the agent processes your prompt, it reads these files first to find relevant information. For best results, use files that are text-heavy and clearly organized.

    Note:
    Use the trash can icon to delete files from this version of the workflow. To delete context files permanently, see Delete an uploaded context file permanently.
  12. Optional: On the Settings tab, you can edit the defaults for the following fields.
    • On failure behavior: Controls what happens when a step fails.
      • System default: Uses the system-wide setting configured by an administrator.
      • Continue on failure: The step is marked as failed and the workflow continues to the next step.
      • Stop workflow or branch on failure: The step fails and the workflow or current branch stops execution.
    • Timeout: The maximum number of seconds the step can run before automation orchestrator marks it as failed with a timeout error. If left unset, the step uses the global default for its step type.
  13. Click Create or Update to save the step and add it to the canvas.

Results

  1. Run the workflow manually or wait for the configured trigger to start it.
  2. Check the status in the builder view or the workflow run view. A green checkmark indicates successful completion. If the agent fails, the status shows as failed and the error details are available in the step's Output pane.
  3. View the output object, which downstream steps can reference later. You can also click the Agent steps tab to review the agent's reasoning, tool calls, and final answer.
  4. If you also configured an Approval step for human verification, click Approvals icon (Approvals icon) in the left navigation and confirm that a new pending request displays with the name you configured.