Microsoft and SJTU open-source Argus: research agent ran 1,548 hours, solved 20-year math problem
xiaohu · x · 2026-09-16
Microsoft and Shanghai Jiao Tong University released and open-sourced Argus, a general agent inference runtime for long-horizon research tasks. Key numbers: one run lasted 1,548 hours across 27 research campaigns, requesting human intervention only once every 40.7 hours on average, with a duty cycle of 95.1%–98.7% — and it solved a 20-year-old open math problem along the way.
The authors argue most agent systems stall once humans step away: without dense reward feedback, agents become sluggish. Argus is built on four principles:
- Evidence-driven: every claim must be backed by verifiable evidence
- Self-evolution: agents learn from their own research history
- Multi-agent collaboration: separate roles with independent contexts but shared workspaces to avoid local hill-climbing
- Core-vertical decoupling: a core layer handles permissions, evidence submission, and human boundaries, while verticals define the actual research tasks
More from coding & agent
- One stale date silently broke a WhatsApp reply agent: 19 messages unanswered — petrusenko_max · 2026-11-02
- The 10X engineer is now just a SKILLS.md: 6 workflow rules for AI coding agents — mdancho84 · 2026-11-02
- Auditing my MCP credential broker: the worst hole was the config file — Relevant_Mousse_9117 · 2026-11-02
- AWS's open-source Strands Harness claims 45% cheaper than Claude Code, reigniting harness debate — blaizedsouza · 2026-11-02
- Xiaomi MiMo's CodeSpecWeave trains coding agents on interdependent specs for long-horizon coding — XiaomiMiMo · 2026-11-02
- Sketchy v0.1.6: AI diagrams use 29% fewer output tokens than plain text — Nkgsporty · 2026-11-02