Do Multi-Agents Dream of Electric Screens? Achieving Perfect Accuracy on AndroidWorld Through Task Decomposition

Researchers have officially introduced Minitap, a groundbreaking multi-agent artificial intelligence system that achieved a perfect 100% success rate on the AndroidWorld benchmark in February 2025. This technological milestone, documented in a recently published paper on the arXiv preprint server, marks the first time an autonomous system has fully solved all 116 complex tasks within the evaluation suite, notably surpassing the average human success rate of 80%. Developed to overcome the limitations of previous AI frameworks, Minitap represents a significant leap forward in the capability of mobile agents to navigate and operate smartphone interfaces with total accuracy. The development team’s study highlights severe architectural flaws in traditional single-agent systems that previously prevented them from mastering the benchmark. These issues include 'context pollution,' where reasoning traces become cluttered with irrelevant data, and silent text input failures where the agent incorrectly assumes a command was executed. Furthermore, single-agent models frequently fell into repetitive action loops, repeating the same errors without a mechanism to escape or correct their behavior. Minitap was specifically engineered to mitigate these problems through task decomposition and cognitive separation, ensuring that different agents handle distinct aspects of a command to maintain clarity. Technically, Minitap utilizes a multi-agent structure that separates different cognitive functions to avoid the pitfalls of mixed reasoning. By delegating specific sub-tasks to specialized components, the system prevents the 'forgetting' or 'confusion' typical of large language models when faced with long operational sequences. This approach allows the system to verify its own inputs and break free from circular logic patterns that would otherwise stall a digital assistant. The success of this methodology suggests that complex mobile navigation is better handled by a collaborative ecosystem of specialized agents rather than a monolithic model. The implications for this breakthrough extend far beyond academic benchmarks, signaling a future where digital assistants can manage smartphones with zero-error reliability. As AI agents move closer to full autonomy, the Minitap architecture provides a blueprint for creating more dependable user interfaces across various operating systems. By outperforming humans on standardized digital tasks, this system sets a new standard for the efficiency and precision of mobile AI interaction, potentially transforming how consumers engage with mobile technology in their daily lives.