Microsoft and SJTU open-source Argus, an agent runtime that drives research for days
机器之心 · wechat · 2026-09-07
Microsoft, Shanghai Jiao Tong University and collaborators open-sourced Argus, a general agent reasoning runtime for multi-day autonomous research (arXiv:2608.05144).
Core idea: from Goal-Driven to Evidence-Driven
Current agents automate execution (the Harness) but not the "driving" above it. Argus adds a Driver layer: next steps are decided by accumulated evidence rather than the initial goal, avoiding goal rigidity. The Driver continually answers four questions: is the work done and good enough, what's the best next step, how should lessons change system behavior, and does anything require a human decision?
Architecture
- Long projects become persistent Campaigns split into well-bounded Missions.
- Loop: Manager→Planner→Engineer⇄Reviewer→Manager, with separate per-role contexts and a shared workspace to prevent local hill-climbing and improve token efficiency.
- Can mix harnesses (Pi, Codex, Claude Code, DeepSeek) in a single task.
- Only evidence-gated experience enters the Wiki/Skill store, scoped at Project/Vertical/Global; Core is decoupled from verticals defined by domain experts.
Results
The report covers 27 campaigns over 1,548 wall-clock hours, one human intervention every 40.7 hours on average, and 95.1–98.7% duty cycle. Within a month of open-sourcing, Argus delivered results across AI4System, AI4Science, chip design, AI4Math and AI4AI, and published the first end-to-end math-conjecture-solving logs and full session traces.
More from coding & agent
- AI agent runs 12-hour experiments to autonomously build a fighting game — msg · 2026-09-07
- Vibe coding's hidden benefit: prototype dumb ideas fast, then toss them — generativist · 2026-09-07
- Serving a local Mac model to code from your phone: four workflow problems — former_farmer · 2026-09-07
- Understanding "Prompt Debt" in ~5 minutes: prompts are becoming tech debt — dbreunig · 2026-09-07
- Dev says he has his Jarvis: no longer reads agent replies at all — BLUECOW009 · 2026-09-07
- Median OpenAI researcher spent $0/day on coding agents in February, post reveals — zacharynado · 2026-09-07