Agent S3 Achieves 69.9% on OSWorld, Nearing Human-Level Computer Use
xwang_lk · x · 2026-07-31
Agent S3, a computer-use agent developed by SimularAI, has achieved 69.9% on the OSWorld v1 benchmark, marking a 10% improvement over the previous state-of-the-art and nearing the human baseline of 72%.
The research has been accepted by TMLR. The team attributes this rapid progress to scaling agents correctly via Behavior Best-of-N and simplifying the baseline design. The project remains 100% open-source.
More from coding & agent
- Open-Source AI Agent Validates Startup Ideas in 10 Minutes — tom_doerr · 2026-07-31
- Positron's Jupyter Notebook Editor Goes GA: AI-Powered IDE for Data Science — HamelHusain · 2026-07-31
- How to Keep Chat History Forever in Codex and Claude Code — HanchungLee · 2026-07-31
- Deep Dive into Buzz: Architecture Principles and Home Server Deployment — daniel_mac8 · 2026-07-31
- n8n + Gemini Automates Cold Emails: From Search to Personalized Send — Mohd_Hamid · 2026-07-31
- Open-Source Claude Code Writing System: Turn Fragmented Notes into Multi-Platform Content — tom_doerr · 2026-07-31