At present, despite the fact that Agent technology has already forged ahead to develop execution capabilities like tool invocation, it still leans heavily on continuous human decision-making when tackling long-duration tasks. This reliance brings to the fore issues such as inflexible objectives and sluggish execution, especially in scenarios where dense feedback is lacking. To surmount these hurdles, Microsoft, in collaboration with Shanghai Jiao Tong University and other esteemed institutions, has jointly open-sourced Argus. Argus is a versatile Agent reasoning runtime, purpose-built for long-duration research tasks. Unlike the traditional goal-driven paradigm, Argus embraces an evidence-driven model. Through groundbreaking designs, including multi-role collaboration and the decoupling of core and vertical domains, it facilitates uninterrupted cross-domain research spanning several days, all without the need for dense standardized feedback.

Argus adopts a hierarchical project structure, dividing projects into two distinct levels: Campaign and Mission. It propels tasks forward through a closed-loop collaboration among four key roles—the Manager, Planner, Engineer, and Reviewer. Simultaneously, it accumulates invaluable experience for future reuse. After undergoing a rigorous 1,548-hour testing phase, Argus has proven its mettle, requiring human intervention only once every 40.7 hours on average, with a duty cycle fluctuating between 95.1% and 98.7%. Moreover, Argus has notched up concrete research achievements across multiple domains, including AI4AI, GPU Kernel optimization, and chip design. It marks a significant leap from automated delivery to automated exploration, culminating in fully autonomous research throughout the entire workflow. In doing so, Argus pioneers a novel approach to research organization, ensuring that research endeavors can persist unabated, even when humans step away from the screen.