Microsoft and Shanghai Jiao Tong University-backed Argus goes open source for multi-day autonomous research
Microsoft, Shanghai Jiao Tong University and other institutions have open-sourced Argus, a general-purpose runtime designed for long-horizon research agents, alongside a technical report, code repository, project site and related open repositories. The project focuses on a gap the authors describe in current agent systems: many can execute tasks through a harness, use tools, modify code and run experiments, but they still depend on humans to keep steering when work stretches from minutes into days.
Argus is built around an evidence-driven approach rather than a purely goal-driven one. The system introduces a "Driver" layer above the harness, intended to decide what should happen next based on accumulated evidence, not just an initial target set when information is sparse. Its runtime organizes persistent Campaigns into bounded Missions and uses a multi-role loop of Manager, Planner, Engineer and Reviewer. The report says the architecture can also combine different harnesses, including Pi, Codex, Claude Code and DeepSeek Harness, while separating core runtime controls from domain-specific vertical definitions.
According to the report, Argus covered 27 Campaigns and 1,548 hours of wall-clock time, with one proactive request for human intervention every 40.7 hours on average and a duty cycle of 95.1% to 98.7%. The team said it has applied the system across AI4AI, GPU kernel work, model training, AI4Science, chip design, AI4Math and AI4System tasks.