Start with the last confirmed app state and find the first transition that did not happen. A model saying “done,” a successful tap, or a dismissed spinner does not prove the requested change was saved. Confirm the result in the app’s visible state or in trustworthy app data, then use screenshots, accessibility information, logs, and network evidence to isolate the failure.
This workflow applies across mobile agents, but platform-specific checks below apply only when you use the named platform or integration.
1. Reproduce the failure and define what success looks like
Write down the exact request and the result the user expects—for example, not just “submit the form,” but “the app displays a confirmation and the submitted item appears in the relevant list.” Then record the conditions around the failure:
- App version and build, operating system, device or simulator, and agent or SDK version.
- Account state and any relevant setup or permissions.
- Whether the failure happens every time, and the shortest sequence that still reproduces it.
- The expected and actual screen or state after each step.
Break a long task into short steps, each with an observable result. For Firebase App Testing agent, the documentation specifically recommends shorter steps for complex tasks and a final-screen assertion that describes what should be visible. See Firebase’s AI-assisted app testing guidance.
#1 Best Overall
2. Find the first divergence between the screen and the agent’s understanding
Inspect the last screenshot or accessibility snapshot before the task went wrong. Compare what the agent selected with what was actually on screen. Check whether the intended control was visible, labeled, exposed to accessibility, and actionable—and whether the screen changed after the action.
Look for an intervening surface that could have redirected input or hidden the control:
- An in-app modal or system permission prompt.
- The keyboard covering a button or changing the layout.
- A loading state, unexpected navigation, or a screen that has not finished updating.
- A scrollable page where the target is off-screen or the need to scroll is not visually obvious.
For Firebase App Testing agent, enable Show agent view to inspect the elements detected through accessibility information, and review the video, logs, and other run artifacts. Firebase documents potential scrolling problems and says the preview cannot navigate an app that uses FLAG_SECURE; the screen appears blank to the test agent. These are limits of that Firebase preview, not a general rule for mobile agents. Details: Firebase AI testing documentation and Firebase’s documented limitations.
Rank #2
- Ideal for Gifting
- Ideal for a bookworm
- Compact for travelling
On iOS, XCTest with XCUIAutomation can query UI elements, inspect snapshots, wait for app state, and assert element properties. XCTest can also monitor unrelated UI interruptions that block controls under test. See Apple’s UI testing documentation.
3. Check whether the agent has the right capability and permission
First identify how the agent is meant to act: by recognizing and controlling the screen, or by calling a declared app function or API. A UI action can fail because the target is inaccessible or obscured; a capability call can fail because it is unsupported, unavailable, or unauthorized.
Android AppFunctions
If the integration uses Android AppFunctions, confirm that the device runs Android 16 or higher, that the relevant function is available and enabled, and that the caller has EXECUTE_APP_FUNCTIONS permission to discover and execute functions. Android describes AppFunctions as experimental-preview technology, so check current availability for the specific workflow. See the Android Developers AppFunctions overview.
Android AccessibilityService
If the agent relies on AccessibilityService, check that the service is enabled and that its use complies with Google Play policy. Google Play requires a narrow, clearly understood purpose for general automation and prohibits using AccessibilityService to let general apps autonomously initiate, plan, and execute actions or decisions. Apps that do not qualify as accessibility tools also need prominent in-app disclosure and affirmative consent. Policy compliance is a requirement, not a workaround for a failed action. See Google Play’s AccessibilityService policy.
4. Rule out mobile SDK and runtime configuration failures
When a mobile SDK connects the app to an agent, confirm that its native module is linked into the build, initialization completed before any bridge or launch calls, required configuration is present, and the intended service or agent deployment is active. Check whether the integration requires the app to be in the foreground.
Quick wins for a faster PC:
Repair Windows errors before they cause bigger problemsFix Now →Scan for outdated or missing drivers - takes under a minuteDriver Scan →Clear out junk files and repair common Windows errorsFree Scan →Salesforce’s Agentforce React Native SDK illustrates how specific these checks can be: its guidance covers installing iOS pods and opening the workspace, registering the Android package, using the correct case-sensitive deployment name, awaiting configure() before launch, and launching from a foreground button handler rather than a background task. Apply those details only to the Salesforce SDK; consult the matching setup and troubleshooting instructions for other integrations. See Salesforce Agentforce React Native troubleshooting.
Rank #4
5. Test endpoint, authentication, and connectivity with evidence
Verify the exact service endpoint configured in the app, rather than assuming a login URL is the correct API endpoint. Confirm the device can reach that service and capture the SDK logger output, ideally before initialization, so the first configuration or session error is not missed.
For Salesforce Agentforce, the troubleshooting guide notes that an incorrect endpoint can produce a 400 error when a session starts or leave a loading spinner running indefinitely. For Employee Agent authentication, it documents refreshing credentials and, if refresh fails, logging out and back in. These are Salesforce-specific remedies; use the relevant vendor’s guidance for another agent. See Salesforce’s troubleshooting guide.
Use request and response records to distinguish three cases: no request was sent, a request failed, or the request succeeded but the app did not show the expected update. Expo’s agent-device tooling can collect focused logs, available network requests and responses, UI state, screenshots, traces, and performance data. Availability of particular artifacts depends on the workflow. See Expo agent-device documentation. Do not diagnose a network problem from a failed task alone; look for a connection error or request failure.
Do these 3 things before closing this tab:
1Clear out junk files and repair common Windows errors2Fix the driver behind crashes, sound loss and screen glitches3Repair Windows errors before they cause bigger problemsBest Value
- It can be a gift option
- Comes with secure packaging
- Helpful in various ways
6. Choose evidence that isolates the likely failure layer
Use the evidence closest to the point where the observed state first diverges. These methods are platform- and tool-specific, not interchangeable requirements for every agent.
| Method | Best for isolating | Evidence to inspect | Applicability and limits |
|---|---|---|---|
| Firebase App Testing agent | Task steps, UI perception and actions, and whether the expected final screen appeared. | Agent view, video, logs, run artifacts, and a final-screen assertion. | Firebase’s AI testing preview. Documented constraints include a five-minute run limit, scrolling shortcomings, and inability to navigate FLAG_SECURE apps. See Firebase guidance and known limitations. |
| Android AppFunctions | Whether the declared function path is supported, discoverable, and authorized. | Device Android version, function availability, and caller permission. | Android 16 or higher; experimental-preview technology. See Android Developers. |
| iOS XCTest with XCUIAutomation | UI element queries, state changes, and interruptions blocking controls. | Snapshots, element queries, waits, and assertions. | For iOS UI testing. See Apple XCTest. |
| Expo agent-device | UI state, logs, requests, and reproducibility of a session. | Screenshots, focused logs, available network records, traces, and performance data. | For supported Expo agent-device workflows. See Expo documentation. |
Firebase’s documented five-minute limit applies to its AI-guided testing preview: a test must succeed within five minutes or it ends early as a failure. It is not a general timeout for mobile AI agents. See Firebase’s known limitations.
7. Rerun the shortest case and verify the side effect
- Make one targeted change based on the evidence—for example, expose or relabel the control, handle an interruption, correct SDK initialization, or fix a confirmed endpoint or credential problem.
- Rerun the shortest reproducible task under the same recorded conditions.
- Check the visible result or reliable app state that proves the requested change took place; do not treat the agent’s completion message or successful tap as proof.
- Keep a regression check for the confirmed outcome. On iOS, XCUIAutomation can query elements, wait for app state, and assert expected properties. Expo agent-device can capture screenshots and record or replay a deterministic session.
Firebase can replay successful action sequences later, but a final assertion still matters; if replay fails, Firebase may fall back to AI actions. See Firebase AI testing documentation and Expo agent-device documentation.
Quick Recap
Product prices and availability are accurate as of the date/time indicated and are subject to change. Any price and availability information displayed on Amazon at the time of purchase will apply.
Free tools Windows power users keep installed
One-click scans. No signup required.
