Microsoft's CUAWright: bash-only agent harness beats GUI by 44% on Odysseys
MicrosoftResearch · hf · 2026-10-06
Microsoft Research introduces CUAWright, a minimal 3K-line terminal harness for digital agents that uses bash commands as its sole action interface and a file system as an evolvable space for dynamically creating tools and managing context.
- Against the published GPT-5.5 baseline, it delivers a 33.2% relative improvement in partial reward on OSWorld 2.0 while cutting estimated cost by 37.5%.
- It outperforms GUI-native harnesses by 4.7% on Online-Mind2Web and 44.0% on the long-horizon Odysseys benchmark.
- On CAD tasks (CADGenBench, BenchCAD), it yields 8.1%–41.6% relative gains over other CLI-based harnesses with GPT-5.5.
- Takeaway: digital environments are far more programmable than their GUIs imply, and a minimal terminal-focused harness is the key to better performance and efficiency.
More from coding & agent
- Open-source LCU decouples Codex Computer Use so any AI agent can drive it via MCP — xiaohu · 2026-10-06
- Indie dev's agents auto-produced 53 pregnancy reels at ~$0.03 GPU cost per clip, skill released free — victor_explore · 2026-10-06
- Jev playbook: use a $0.042/M-token judge model to route agents and clean context — blaizedsouza · 2026-10-06
- "I don't code anymore, I just yap and stuff happens" — jaivinwylde · 2026-10-06
- Prompt craft: using Opus 5.5 with JS Canvas to draw a Chinese seal-script seal — dotey · 2026-10-06
- "It's insane that coding is now just audibly speaking at your computer" — KevinNaughtonJr · 2026-10-06