CUAWright: A Minimal Unified Interface for Digital Agents
AuthorsYadong Lu, Theodore Lee, Yifei Li, Lawrence Keunho Jang, Tianci Xue, Yu Su, Huan Sun, Ahmed Hassan Awadallah
AffiliationsOct 2026 · Microsoft Research National University of Singapore The Ohio State University · Carnegie Mellon University
CUAWright argues that giving computer-use agents a simple programmable terminal instead of GUI-specific tools makes them substantially more capable, efficient, and adaptable.
Key results
Gain over the GUI-based approach using the same model.
Gain over the GUI-based approach using the same model.
Compared with 35.1% for Codex CLI.
What the paper found
CUAWright replaces domain-specific GUI harnesses with a roughly 3K-line terminal-first interface: agents act through bash, use terminal output by default, request screenshots only when needed, and treat the filesystem as persistent memory and a place to build reusable tools. Across web, desktop, and CAD tasks, tests with GPT-5.5 and GPT-5.6 Sol show gains over GUI-based approaches; the comparisons also include OpenAI’s published GPT-5.5 harness and Anthropic’s Claude Opus baselines. On OSWorld 2.0, CUAWright reports a 33.2% relative improvement in partial reward and a 37.5% reduction in estimated cost versus the published GPT-5.5 baseline. On Online-Mind2Web and Odysseys, success rates improve by 4.7% and 44.0%, respectively. On WeaveBench, it reaches a 53.5% pass rate, compared with 35.1% for Codex CLI. For CAD reconstruction on BenchCAD, it achieves 79.0% mean IoU, compared with 55.8% for the published GPT-5.5 result. The results suggest that direct programmatic access and agent-created tools can make digital workflows more efficient than relying primarily on screenshots and mouse-and-keyboard actions.
Original abstract
The prevailing approach to computer-use agents couples a model with a domain-specific harness: a browser or desktop environment equipped with human engineered tools that are fixed before task execution. As models' coding capabilities improve, the GUI native and static harness prevents them from direct programmatic operation on system state, as well as flexible construction of tools. To this end, we introduce CUAWright, a minimal terminal harness of roughly 3K lines of code that uses bash commands as its sole action interface, and a file system as its evolvable space for dynamically creating tools and managing the context. We conduct comprehensive experiments across a wide range of digital tasks, and demonstrate that by giving the agent a minimal, programmable interface, it achieves substantially stronger results compared to their GUI or hybrid CLI interface across a wide range of tasks. On OSWorld 2.0, CUAWright delivers a 33.2% relative improvement in partial reward while reducing estimated cost by 37.5% compared with the published GPT-5.5 baseline. On Online-Mind2Web and the long horizon Odysseys benchmark, CUAWright substantially outperforms GUI native harness by 4.7% and 44.0% in success rate, respectively. Furthermore, we found the gains extend to CAD applications that require accurate visual understanding and CLI interaction: on CADGenBench and BenchCAD, our unified harness yields 8.1%-41.6% relative improvements over other CLI based harnesses with GPT-5.5. Together, these results suggest digital environments are far more programmable than their GUI interfaces imply, and a minimal terminal-focused harness is the key for better performance and efficiency.
Read the original paperMore in AI Agents
Browse all 62 papers →MAI-Gym: An Open All-in-One Platform for GUI Agent Evaluation and RL Training across Mobile and Computer Environments
Panrong Tong, Quyu Kong, Pengxiang Li, Tianyu Xia, Hanzhang Zhou, Lei Wang, Chenglin Cai, Long Li, Long Chen, Jianan Zhang, Zhengwen Zeng, Xu Zhang, Gongjie Zhang, Yue Wang, Steven Hoi
MAI-Gym is an open platform for training and testing GUI agents at scale, with reported gains that transfer from simulated apps to a physical phone.
Stateless Language Agents: Scaling Long-Horizon Automated Research
Qizheng Zhang, Changxiu Ji, Isaac Sun, Yuetai Li, Shubhangi Upasani, Sherry Ruan, Boyuan Ma, Fenglu Hong, Vamsidhar Kamanuru, Yoonho Lee, Yuzhen Mao, Genghan Zhang, Rulin Shao, Qiuyang Mang, Andy Dimnaku, Changran Hu, Radha Poovendran, Kunle Olukotun
This paper shows how LLM research agents can make longer-term progress by keeping shared research state outside their conversations and assigning focused experiments.
DAEDALUS: Bootstrapping Agent Memory from Self-Generated Tasks
Antoine Edy, Max Conti, Victor Xing, Marc-Antoine Allard, Nawfal Benhamdane, Gautier Viaud
DAEDALUS lets AI agents invent practice tasks, learn from their failures, and build reusable memories that improve performance in new environments.