Action Center

DELIVERY:

Last updated May 15, 2025

Creating agent escalations with Action apps

Terminology:

Agents: Capabilities powered by technologies like large language models (LLMs), machine learning, and traditional enterprise automation, agents are designed to operate in dynamic, non-deterministic environments. They can plan, act, learn, and adapt—making them ideal for processes that require judgment, flexibility, and contextual awareness.
Unlike deterministic systems such as RPA robots, which follow structured logic and fixed rules, agents use a probabilistic approach to make decisions based on patterns and real-time data. This makes agents highly suited for automating unstructured, exception-heavy workflows where conditions and outcomes vary.
Action apps: A type of app designed in UiPath Apps or directly in Studio Web, allowing you to create forms or UI interactions with specific options for tasks requiring human intervention (Action Center tasks), such as approvals. You can define the input and output properties of action apps by adding action properties in the action schema.
Note: You can only use published and deployed apps to configure agent escalations.

There could be situations where agents face challenges or require human intervention to assess, confirm or provide essential details. This might be a deliberate action, where certain steps are always assigned to a human before transitioning back to an agent. Some examples where escalations may be necessary are the following: A tool call can fail unexpectedly or return data of an unexpected structure, they might need help discerning between multiple customers with the same name in a third-party app, or they might have low confidence in an action. There might also be organizational reasons that they aren’t trusted to make certain types of decisions or take certain actions.

In all of these cases, agents need to be able to escalate to a human and:

Surface all relevant details in an effective, straightforward, and customizable way.
Allow the Human-In-The-Loop (HITL) to provide the relevant decision or information, whether this is via a simple Approval/Denial decision for a case or via providing several business details - emails, order numbers, or references to accounts in external platforms.

Agents might even need for one of their tools to be fixed before they can continue operation on a business case. You can use the Guardrails tool to automatically manage irregular tool calling. For more information, visit Guardrails. When creating escalations for guardrails, you can use the following UiPath Marketplace template: Agent Tool Guardrails Escalation App.

What are escalations?

Escalations are tools used to involve users in an agent's running process. Escalations are used to evaluate or validate an agent's decisions or supply input that the agent might not have. Therefore, humans still have a vital role in maintaining an agent's health and proficiency. You can set conditions for specific escalations through the following configurations:

System or User prompt: Define the agent's objective and clarify when to use a specific escalation.
Escalation - Under the escalation component, select an Action app designed for a particular circumstance. By giving the Action app a unique name and description, the agent can better understand which escalation it should use and when, along with the instructions from the system and user prompt. You can add multiple escalations to a single agent. You can assign multiple escalations to one agent and add a description to each input, enabling the agent to send appropriate values to the Action app. All these settings significantly impact the agent's effectiveness and overall health score. . For more information on an agent's score, visit Agents score.

Important: Escalations are non-deterministic, meaning the agent uses provided instructions to decide when and what to escalate. If you need a consistent and deterministic escalation, it is recommended to use an Action app outside the agent. You would instead end the agent and use its outputs in an App action. Make sure you consider these recommendations, according to your use cases.

1. Create an Action app

First you need to create and configure a new Action app. The steps in this procedure show how to create an app that approves search results. You can adapt these steps to your particular use cases.

Note: You can configure agent escalations only using Action apps. We recommend you use Action apps created in Studio Web, rather than Action apps created in UiPath Apps.

Create an Action app from scratch

Go to Automation Cloud™ and select the Studio Web service.
From the Automations tab, select New project, and choose the App project type.
Select a preferred template, style, and the desired controls. For the purposes of this tutorial, use the Form D template.
Read more about designing your app.
Select App name, then, under the General tab, select App name to open the expression editor.
Give the new app a name: "Tool Approval App" and save.
From the left navigation panel, select FormWrapper and remove the controls that you are not going to use in this example: FirstName, LastName, Email, Phone, Street, State, Zip, and Country.
Select the project explorer button on the top-left of the page, then select the + add button, and choose the Action option.
From the Toolbox button on the top-left of the page, navigate to the Display tab, and add a label to the app.

Now you need to turn this App into an Action app, otherwise you would not be able to use this app to configure an agent escalation.
Select the project explorer button on the top-left of the page, then select the + add button, and choose the Action option. The Action properties dialog displays.
In the Action properties dialog, add webSearchResults in the Input Properties and make it Mandatory by checking the related box.
The action schema's input properties are derived from the agent context and presented to the user. They are likely to be unique for each agent run. Note that, unlike Input/Output properties, you cannot change these values during review. The Input/Output properties are derived from the agent context and you can modify them. The agent also documents any changes returned to it. Output properties are additional values that users can enter to be sent back to the agent.
Create two outcomes: Approve and Deny.

The outcome is an event that, along with the Input/Output properties, is sent back to the agent. Providing a meaningful outcome guides the agent's next steps, which might involve ending its run due to failure. Therefore, testing is crucial.
Return to the form and name the label you previously created in the expression editor with "webSearchResults for approval is: "+ActionProperties.webSearchResults. Select Save.
Change the names of your Submit and Cancel buttons into Approve and Deny.
Select the Approve button and go to the Events tab on the right.
Add a "Clicked on" automation by selecting the Edit automation button.
Add a Submit Action activity.
Read more about adding activities in Studio Web.
In the Outcome section, choose the Approve value.
Select the Deny button and go to the Events tab on the right.
Add a "Clicked on" automation by selecting the Edit automation button.
Add a Submit Action activity.
Read more about adding activities in Studio Web.
In the Outcome section, choose the Deny value.

Create an Action app using Autopilot^TM

Go to Automation Cloud™ and select the Studio Web service.
From the Automations tab, select New project, and choose the App project type.
Select the Blank Page template.
To leverage Autopilot capabilities in creating the Action app, you need to have a blank app.
From the Project Explorer, right-click the app project, and from the Add to the app dropdown list, select Action.
This automatically opens the Action schema inside the app.
In the With Autopilot section, input the instructions for Autopilot to generate the action schema.
For example, you can enter: "Create an action schema, were an input argument is Age, an in/out argument is salary, and an out argument is interest rate".
Select Generate Action to generate the action schema.
Select Generate page from schema to generate an app configuration based on the previously generated action schema.
Refresh the Studio Web project.
Your Action app is now ready to be published and deployed.

Publish and deploy the Action app

To make this Action App available in the organization, you need to publish and deploy it:

Select Publish.
Choose the name "EscalationTestApp", and select Publish again.
Depending on where you are creating the Action app, deploy the app as follows:
- If you created the app in an App project, go to Orchestrator and deploy the app.
- If you created the app as part of a solution in Studio Web, navigate to the Solutions tab in Orchestrator, and deploy the app.
Make sure it runs on a serverless runtime.
After deploying the app, select the Edit the process details button.
Select Next.
In the Cloud Robot - Serverless section, choose the Medium option. This is the option we recommend for best performance with Web Apps.
Select Update.

2. Create an agent

Create a new Agent following this agent building procedure.
In the System prompt field input the following: Use the web search tool to do a search on the provided search item. You must then raise an escalation to confirm that the search result is satisfactory before returning your final output, a simple one-liner from the web search.
In the User prompt field, input the following: Do a search on term:.
In the left-side part of your project, select Open Data Manager, and on the Input tab, select Add property.
Select Edit Name and enter searchTerm. Select Apply to save the changes.
In the Description field, input the search term provided from the user.
On the Output tab, select Add property, and set the name as Output and the description as natural language summary of the approved search results.
Select Required to make this property required.
Figure 1. The User and System prompt of the Agent
Go back to the User Prompt, and add the previously created property using double curly brackets {{ }}.
In this example, enter double opening curly brackets {{, and then select searchTerm.
In the Tools section, select Add tool, and go to the Activities tab.
Search and select the Web Search activity.
Open the Properties panel for this activity, and in the UiPath GenAI Activities Connection dropdown list, select an existing connection.
Alternatively, select Define new to create a new connection, or select Open connections to navigate to Integration Service.
From the Search engine field, select a GoogleCustomSearch.
In the Search description section, from the tune menu, select Argument, and select the previously created argument, which should contain the topic you want to search the web for.
The agent also uses the description of the argument, provided in the Data Manager, as part of the user prompt.

Figure 2. The Properties panel of the Web Search activity
In the Escalations section, select Add escalation:
1. Select the previously created and deployed Action app in the App dropdown list.
  In this example, select EscalationTestApp.
2. In the Prompt field, input Use this to confirm the search results is satisfactory with the assignee.
3. Expand the Inputs section to view the fields of the Action app.
Figure 3. Configuring the agent escalation
Run the agent.
From the agent's Traces tab, select the URL of the task that is pending with the previously configured assignee.
1. Review the escalation task awaiting you and approve it if everything is in order.
Note: The agent will be in a suspended state until the escalation is resolved by the assigned user in Action Center.
Go back to the Agents workspace, where you should see the Agent executing the task and providing the expected output.