9.2 Agentic UI Automation: ScreenPlay, Semantic Activities & Healing Agent
Key Takeaways
ScreenPlay (formerly UI Agent, class UiPath.Semantic.Activities.NUITask) performs a UI task from a natural-language prompt and must sit inside Use Application/Browser.
ScreenPlay guardrails include Max number of steps, variable security against prompt injection, and trace attachment to jobs.
Semantic selectors act as a fallback or primary targeting method, and semantic activities (Fill Form, Extract Form Data, Update UI Element) handle forms by meaning.
Healing Agent gives recommendations and self-healing for UI Automation Modern activities in unattended, attended, and test runs, with Test Heals and RPA Heals tracked separately.
Healing Agent is enabled through AI Trust Layer governance and Orchestrator settings, and each activity has a Healing Agent mode.
9.2 Agentic UI Automation: ScreenPlay, Semantic Activities & Healing Agent
Core Concept: The January 2026 exam description adds "UiTask/UiAct — AGENTIC (healing scenarios, testing, attended scenarios)" under Advanced UI automation. UiPath's documentation does not define "UiAct" as a product name. What it documents is a family of agentic UI features: ScreenPlay (implemented as the activity class
UiPath.Semantic.Activities.NUITask), semantic activities and semantic selectors, and Healing Agent, which repairs failing UI automations. Learn what each one does, where it runs, and how it is governed.
ScreenPlay: Describe the Task, Let the Agent Act
ScreenPlay (earlier called UI Agent) is UiPath's agent for driving an application's interface from a natural-language instruction. It "interacts with applications much like a human would", navigating screens and adapting to changes instead of following fixed selectors.
| Property | What it does |
|---|---|
| Task | The prompt describing the UI task. It can include variables, attached files, and images captured from the screen, and you can view the last execution trace. |
| Model | The large action model that plans and reasons. UiPath offers several Screen Agent models with different speed, cost, and capability trade-offs. |
| Max number of steps | A guardrail that caps how many actions the agent may take, which prevents endless agentic loops. |
| Job attachment mode | When to attach the .html execution trace to the Orchestrator job: On failure, Always, or Never. |
| Use DOM when available | Sends DOM data to UiPath's models where it can be extracted. Turn it off if DOM-based targeting leads to wrong coordinates. |
| Disable variable security | Variable security stops prompt instructions from being injected through variable values; disable it only for a confirmed false positive. |
| Input mode | The same choices as other UI activities: Same as App/Browser, Hardware Events, Chromium API, Simulate, or Window messages. |
Key rules:
- ScreenPlay must be placed inside a Use Application/Browser activity.
- It is available in Windows and cross-platform projects, from the UIAutomation package versions that include it (UiPath's setup guide names 2025.10.20 or newer), with Studio connected to Automation Cloud.
- It is governed through the AI Trust Layer, with protection against prompt injection.
UiPath's own advice is to keep workflows "maximally deterministic and minimally agentic": use ordinary activities for predictable steps and ScreenPlay only where selectors are fragile or real reasoning is needed. ScreenPlay does not remember fallback paths between runs; each run reasons again from the current screen.
Semantic Selectors and Semantic Activities
These features apply AI to individual targets and forms, with tighter limits than ScreenPlay:
- Semantic selectors use semantic understanding of the screen as a targeting method. They work as a fallback, triggered when traditional or Computer Vision targeting fails, or as a primary method for highly dynamic interfaces. UiPath bundles a monthly number of semantic selector runs with each platform tier.
- Semantic activities handle forms: Fill Form, Extract Form Data, and Update UI Element (formerly Set Value). They match fields by meaning, which makes it easy to move data between web forms whose layouts differ.
UiPath calls these "micro agents": bounded, task-specific intelligence with predictable execution.
Healing Agent: Recovering Failing UI Automations
Healing Agent analyzes a UI automation just in time and chooses recovery strategies when it fails. It offers two benefits:
- Recommendations. When a UI automation fails, it can suggest new selectors, smart delays, or code changes to handle unexpected pop-ups. These appear with the job in Orchestrator so a developer can update the workflow.
- Self-healing. It complements Unified Target's own fallback by applying one or more recovery steps in a cascade, for example closing an overlay, adjusting a selector, and waiting for the page to load, so the job keeps running.
| Scenario | How Healing Agent helps |
|---|---|
| Unattended | Background jobs on VMs that fail because of load times, unstable selectors, or pop-ups |
| Attended | Automations running alongside a user, where user actions or UI changes break a step |
| Testing | Healing during Test Manager and Test Cloud executions, counted as Test Heals and licensed separately from RPA Heals |
Important details:
- Healing Agent supports UI Automation Modern activities only, for desktop and web applications, and it can recover automations running in remote environments such as Citrix.
- It is enabled through governance: an AI Trust Layer policy, plus settings at the Orchestrator process or job level. If self-healing is not activated, it only provides debug information and limited recommendations.
- Individual activities have a Healing Agent mode property: Disabled, Inherit job settings, Recommendation only, or Same as App/Browser.
Choosing Among Them
| Need | Best fit |
|---|---|
| A stable, high-volume process with good selectors | Ordinary Modern activities, with Healing Agent as a safety net |
| One field or element whose selector keeps drifting | A semantic selector as a fallback or primary targeting method |
| Moving data between forms with different layouts | Semantic activities such as Fill Form and Extract Form Data |
| A multi-step task in an application that changes often or has no usable selectors | ScreenPlay, with a step limit and trace attachment |
| Test suites that break after UI releases | Healing Agent during test executions (Test Heals) |
A developer adds a ScreenPlay activity directly to a sequence, outside any other container, and the workflow does not validate. What is the fix?
Move ScreenPlay inside a Use Application/Browser activity that targets the application.
Convert the project to Windows-Legacy.
Wrap ScreenPlay in a Parallel activity.
Set Max number of steps to 0.
Which statement about Healing Agent is correct?
It works with both Classic and Modern UI Automation activities.
It replaces Unified Target, so selectors are no longer recorded.
It can only recommend fixes and never changes a running job.
It supports Modern UI Automation activities, can self-heal or only recommend depending on configuration, and counts test executions as Test Heals.
A team needs to copy customer details between two web forms whose field layouts differ and change often. Which agentic feature fits best while keeping execution bounded and predictable?
A ScreenPlay activity with an unlimited number of steps.
Semantic activities such as Extract Form Data and Fill Form.
Healing Agent in Recommendation only mode.
A Hardware Events Click activity with Computer Vision disabled.
Sections you finish are checked off in the contents.