Update time

Our relentless pursuit of mobile AI excellence has paid off. After intensive optimization cycles, we've not only reclaimed our position as the industry leader but pushed the boundaries even further. With a remarkable 77.59% success rate on the Android World benchmark, we're setting new standards for what mobile AI agents can achieve.

This breakthrough wasn't achieved through incremental improvements, it required fundamental reimagining of how AI agents operate on mobile platforms. Here's how we did it.

Tools Strict Validation

We implemented rigorous validation layers that ensure every tool interaction is precise and purposeful. By enforcing strict parameter validation and contextual appropriateness checks, we eliminated the noise that was degrading our agents' decision-making capabilities. This foundational change alone contributed to a significant boost in reliability and execution accuracy.

Cortex Enhanced Reasoning

We developed a revolutionary "reasoning on top of reasoning" architecture where specialized critic agents continuously evaluate and refine the actions of executing agents. This meta-cognitive approach, built on LangGraph's robust framework, creates a dynamic feedback loop that catches potential errors before they cascade into failures. Our cortex system doesn't just think, it thinks about its thinking.

Performance Breakthroughs

🚀 Advanced Task Planning Logic

Our agents now identify and execute parallel subgoals simultaneously. Instead of sequential step-by-step execution, they recognize opportunities for concurrent actions, reducing overall task completion time while maintaining precision.

⌨️ Refined Text Input Tools

We streamlined text field interactions by eliminating redundant focus operations and optimizing input sequences. This change removed significant overhead and improved user experience consistency across different mobile interfaces.

🧠 Intelligent Decision Agents

Enhanced guidance systems now provide non-executing agents with complete tool availability awareness, enabling more informed decision-making and reducing unnecessary exploration cycles.

🔧 Universal Provider Support

Responding to community demand, we've implemented OpenAI-compatible provider support. Whether you prefer local LLMs or unified provider ecosystems, our platform adapts seamlessly to your infrastructure preferences.

What's Next: Continuous Improvement

Our next focus areas include advanced observability features and revolutionary screenshot analysis capabilities that will give our agents unprecedented UI understanding. We're building the future of mobile AI interaction, one breakthrough at a time.

Our next major milestone? Breaking the human benchmark performance barrier at 80%, an important achievement that would push mobile AI forward significantly. But that's just our next step. We're not just competing with other AI systems or aiming to match human performance, we're building toward capabilities that go beyond.

77.59% is a milestone, 80% is our next target. More to come.

Detailed Task Results

Complete breakdown of all 116 benchmark tasks with transparent trace data.

View Detailed Results