Unified Agent Manages Cross-Device AI Interactions.

As AI agents become more sophisticated, their ability to operate seamlessly across a user's various devices and over extended periods is becoming increasingly important. However, current agent systems are not adequately designed for this multi-device, multi-temporal interaction paradigm. Traditional single agents often lack effective state management for observations scattered across different devices and moments, while multi-agent systems, though coordinated, fail to maintain the compact, carried state necessary for a continuous cross-device user request. This research argues for a new approach where the agent maintains an effectively designed state that consolidates engagement evidence, stated facts, and ongoing requests into a concise, action-ready form. This state then informs the agent's actions based on current observations. To validate this principle, a new benchmark for user-agent interaction across devices and time was constructed. The proposed system, Unified Agent, embodies this stateful design. It carries interaction evidence across devices and moments, using it in conjunction with current observations to act intelligently. In evaluations, Unified Agent significantly outperformed adaptations of four published designs. Its advantage remained robust across variations in multimodal large language model (MLLM) families, capabilities, and reasoning efforts, demonstrating the inherent strength of its state-design approach. The code and data will be publicly available.