# 🏆 FinalRun achieves the New #1 SOTA Mobile Agent Eclipsing Benchmarks at 97.4%

We are thrilled to announce [**FinalRun**](https://finalrun.app), a revolutionary Android UI agent that has officially secured the #1 spot on the leaderboard with an unprecedented **97.4% Success Rate**. Not only does FinalRun achieve the highest accuracy ever recorded, but it is fundamentally engineered to be **faster and significantly cheaper** than any other agent currently on the benchmark list.

By rethinking how AI agents perceive UI hierarchies, handle real-time states, and process errors, FinalRun sets a new gold standard for autonomous mobile execution.

* * *

![](https://cdn.hashnode.com/uploads/covers/687a10136bb0eb22986499e2/55d40c4f-fb3d-4006-a1f6-a74588cdf8a2.png align="center")

## The Race for the Ultimate UI Agent

Building an AI agent that can reliably navigate the dynamic, asynchronous environment of a mobile operating system is notoriously difficult. Traditional agents stumble on UI animations, hallucinate successful actions, and incur massive inference costs due to bloated observation spaces.

With **FinalRun**, we completely reimagined the perception-action loop. We focused not just on *reasoning capability*, but on *execution reliability* and *cost efficiency*. The result? A staggering 97.4% task completion rate that operates faster and cheaper than the previous state-of-the-art.

Here are the four core breakthroughs that power FinalRun:

### 1\. ⚡ Optimized Hierarchy

The biggest bottleneck for mobile agents is the massive context window required to read Android UI trees (XML). FinalRun utilizes a highly **Optimized UI Hierarchy**, aggressively pruning irrelevant nodes.

*   **Cheaper & Smarter:** By pairing this lightweight hierarchy with **Gemini Flash** and **Flash Lite**, we drastically reduced token consumption. The cost per task is a fraction of what competing agents require.
    
*   **Blazing Fast:** Powered by Gemini Flash and Gemini Flash light
    
*   **Image Stabilization Check:** Speed means nothing if the agent clicks while the screen is still loading. FinalRun introduces a deterministic stabilization check—requiring **at least 3 consecutive stable image frames**—before concluding the UI has settled. This entirely eliminates misclicks caused by mid-animation transitions.
    

### 2\. 📝 Native Text Editing Mastery

Manipulating text fields is a common failure point for MLLMs, which often try to backspace infinitely or struggle with cursor placement. FinalRun features a dedicated **Text Editing Agent** that interacts with the OS the same way a human power-user does. When specific text needs to be deleted, the agent executes a **Long Press** to trigger the Android system’s native context menu, utilizing default options to **"Select All"** and **"Delete"**. This native approach guarantees 100% precision in text manipulation, avoiding the erratic behavior seen in previous SOTA agents.

### 3\. 🔄 Closed-Loop Failure Feedback

A major flaw in traditional UI agents is the "blind assumption of success"—the agent issues a click command and blindly assumes the task moved forward. FinalRun introduces a strict **Failure Feedback Loop**. If a grounding action (e.g., attempting to tap a coordinate or element) fails at the system level, this failure is immediately fed back into the agent's context. The agent knows exactly *why* the action failed, preventing it from hallucinating task completion and allowing it to instantly recalculate a new path to success.

### 4\. 📸 Just-In-Time (JIT) Grounding

Mobile UIs are highly volatile; pop-ups, notifications, or network-delayed renders can change the screen state in milliseconds. Traditional agents evaluate a screenshot, spend seconds "thinking," and then execute a click on a screen that may have already changed. FinalRun solves this race condition by **capturing a fresh screenshot immediately before performing any grounding action**. This ensures that all coordinates and actions are executed against the *absolute latest* UI state. This Just-In-Time perception captures errors faster, prevents catastrophic misclicks, and ensures a truly real-time interaction loop.

* * *

## The Future of Mobile Automation is Here

Achieving 97.4% on complex mobile benchmarks proves that reliable, production-ready mobile AI is no longer a theoretical concept. By combining the raw speed of Gemini Flash, an optimized UI hierarchy, and robust, human-like execution safeguards, FinalRun is unequivocally the fastest, cheapest, and most capable Android agent in the world.

*Welcome to the new SOTA.*

We’re open-sourcing the agent soon

[Notify me when it's open source →](https://docs.google.com/forms/d/1EwHjqK6t1pBQgsKWih1Z_hQqP837elhl7hVbUGLHqV0)