Agentic RPA Action

Overview

The RUN AGENTIC RPA action lets you set up a robotic task using natural language. Instead of configuring each step manually, you write a plain language prompt describing the goal, and an AI model determines and executes the steps needed to complete it.

This page describes how to configure the RUN AGENTIC RPA action, interpret its output, and install the RPA agent required to run Agentic RPA robotic tasks.

The RUN AGENTIC RPA action appears in the RPA actions palette and integrates seamlessly with existing RPA actions.

Run Agentic RPA

The RUN AGENTIC RPA action uses an AI model to complete a robotic task described in natural language. You provide a prompt, pick a model, and optionally, include an input variable and credentials, and the action returns text describing the outcome.

You can configure CONFIGURATION and OUTPUTS parameters for this action.

Add the Run Agentic RPA action to a robotic task

This section describes the steps to add the RUN AGENTIC RPA action to your robotic task workflow.

To add the RUN AGENTIC RPA action to a robotic task:

1. Drag and drop the RUN AGENTIC RPA action from the RPA actions palette into a section or action group.
2. In the CONFIGURATION area of the pane, configure the parameters.

3. In the OUTPUTS area of the pane, configure the Operator and Target parameters to store the result.

4. In the AFTER COMPLETION area of the action pane, configure the following parameters:

   • Wait before executing next action (seconds): Enter the number of seconds the robotic task waits before it continues to the next action in the robotic task definition.
   • Take a screenshot: If you select this option, the robotic task takes a screenshot of the host machine after the action completes. If the action failed, the screenshot is taken immediately before the exception is thrown and cleanup is started. If the action completed successfully, the screenshot is taken immediately after the action is performed.
5. Click SAVE CHANGES in the page toolbar.

Configuration

In the CONFIGURATION area of the ACTION CONFIGURATION pane, configure the following parameters:

Output

When a RUN AGENTIC RPA action finishes, the action saves output data (in JSON format) to the target variable you configure in the OUTPUTS area. The output contains the following fields:

A successful action returns output similar to the following:

1
2
3
4
{
  "status": "success",
  "result": "Appian's web page (appian.com) is now open and visible in Firefox."
}

A failed action returns output similar to the following:

1
2
3
4
{
  "status": "failure",
  "result": "Opera browser is not present on the desktop or installed on this system, so Appian's web page cannot be opened in Opera."
}

An aborted action returns output similar to the following:

1
2
3
4
5
{
  "status": "abort",
  "result": "",
  "error_details": "The action was stopped because the model repeated the same step 10 times and appears to be stuck."
}

In the OUTPUTS area of the ACTION CONFIGURATION pane, configure the following parameters to store this result:

• Operator: Select an option to store the result as a single-value variable (is stored as) or add the result as a multiple-value variable (is appended to).
• Target: Enter the name of the robotic task variable where you want to store or append the result. If your variable is type CDT, you can specify which field of a CDT to store the result into using dot notation.

Usage considerations

To avoid unnecessary AI action usage, the RUN AGENTIC RPA action stops itself early under the following conditions, and the action ends with an "abort" status:

• The action reaches the configured number of Maximum iterations.
• The action reaches 100 iterations without resolving the request. Appian considers the prompt too vague to resolve accurately. Write specific, well-scoped prompts to avoid hitting this limit.
• The AI model repeats the same step more than 10 times within the last 20 steps. Appian considers the model stuck.

AI action logs

Appian records AI action usage from the RUN AGENTIC RPA action in two places:

• The Execution Log of the robotic task, at the TRACE level. Each entry lists the AI actions used for that response and the running total for the action. For example, an entry might read AI actions - 36.0 | Total - 222.0.
• An AGENTIC_AI_USAGE event is recorded at the end of each RUN AGENTIC RPA action, containing the execution ID and the total AI actions used.

Best practices for agentic RPA prompts

When creating a prompt for an Agentic RPA action, structure your instructions into five distinct sections to ensure reliable, predictable, and safe execution.

1. Objective & Global Rules Define the ultimate goal of the automation and the global rules of engagement.

• State the core task ("Create a user in the CRM" or "Log in to the SAP application").
• Set the autonomy level / Anti-looping ("Complete the process autonomously" or "Do not retry any action under any circumstance").
• Define the global failure state ("If any validation fails, stop processing immediately").

2. Inputs & Context Tell the agent exactly what data it has access to and where it comes from.

• List the specific variables or data maps provided (username, password, InvoiceData).
• Specify credentials or authentication requirements.

3. Step-by-Step Execution & Guardrails Break the process down into chronological steps. For every action, provide:

• The Anchor: For web, use URLs and menus. For desktop/legacy apps, use visual anchors like window titles ("Wait for a window titled 'SAP'") and visual descriptions ("green tick button").
• The Action: What to click, navigate to, or type.
• Fuzzy Logic / Tolerance: Explicitly state how strictly to compare data. If minor typos, capitalization, or currency symbols can be ignored, say so (treat "1250" and "$1,250.00" as matches).
• Alternative Paths: If the primary action fails, define what the agent should try next ("If the default address doesn't match, click the edit icon and check the list").
• Negative Constraints: Tell the agent what not to touch ("Do not modify the secondary contact").
• Verification & Failure: End every step by asking the agent to verify the action succeeded, and instruct it to fail immediately if it didn't.

4. The "Definition of Done" (Checklist) Provide a summary checklist of every milestone that must be achieved for the whole process to be considered successful. This forces the LLM to review its own work before finishing.

• Example: "Return success only if: Login completed, Data was entered, Save was confirmed, and no errors were displayed."
• Example: “Consider the login successful only if 'SAP Easy Access' is visible on the screen."

5. Strict Output Formatting Eliminate the AI's conversational tendencies so your RPA platform can parse the result.

• Specify the exact expected output (true/false, a specific JSON schema, or a specific ID number).
• Add a strict prohibition on extra text: "Do not return explanations, status messages, or any other text."

Example prompt

1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
Objective & Global Rules:

Create a new employee profile in the HR system using the provided input map. Complete autonomously. If any required element is missing, validation fails, or an error occurs, stop processing immediately and return false.

Inputs & Context:

Use the `EmployeeData` map, containing: First Name, Last Name, Personal Email, Department, and Start Date. Use the provided standard system credentials for login.

Step-by-Step Execution & Guardrails:

- Navigate to `https://hr.example.com/login` and enter credentials. If login fails, return false.  
- Navigate to "Directory" > "Add New Employee".  
- Enter First Name, Last Name, and Personal Email exactly as provided.  
- Select the Department. Treat minor formatting variations ("IT", "I.T.", "Information Technology") as matches.  
- Enter the Start Date in the `MM/DD/YYYY` format expected by the system.  
- Do not assign any user roles or system permissions; skip the "Access Configuration" tab entirely.  
- Click "Save Profile" and wait for the confirmation banner to load. If the confirmation does not appear, return false.

Definition of Done:

Return Success only if all of the following occurred successfully:

1. Login completed.  
2. Navigation to the "Add New Employee" screen succeeded.  
3. Name, Email, Department, and Start Date were entered correctly.  
4. Access and permission configurations were bypassed.  
5. The profile was saved and the visual confirmation banner was verified.

Strict Output Formatting:

Return only `Success` or `Failure`. Do not return explanations, status messages, error descriptions, or any other text.

Feedback