Microsoft Unveils Evidence-Driven AI System for Long-Horizon Research Tasks
Microsoft and Shanghai Jiao Tong University have collaborated to develop Argus, an evidence-driven AI system designed for long-horizon research tasks. The system has been open-sourced and has demonstrated impressive results in various domains, including AI4AI, GPU Kernel, model training, AI4Science, chip design, AI4math, and AI4System.
Argus solves a 20-year-old mathematical problem by automating the 'steering' logic above execution. The system enables Agents to expand to multi-domain continuous research for several days without dense standard feedback. It consists of four roles: Manager, Planner, Engineer, and Reviewer, which collaborate to prevent the Agent from becoming a local hill-climber.
The Argus runtime is designed to be customizable and generalizable, allowing different harnesses like Pi, Codex, Claude Code, and DeepSeek Harness to be enabled simultaneously in one task. The system stores a complete set of artifacts and only stores experience that meets the evidence threshold, which can be reused for subsequent tasks.
After 1548 hours of operation, Argus has produced outstanding contributions in infrastructure, materials, chips, mathematics, and AI system research. It has also published the first end-to-end screening log for mathematical conjecture solving and made its operation track sessions publicly available.