CUAWright: A Minimal Unified Interface for Digital Agents

CUAWright. A Minimal Unified Interface for Digital Agents. By Yadong Lu and seven others, from Microsoft Research, the National University of Singapore, and two other institutions. Published on the second of October. Its code repository is worked out: about seventy-five source files, with a readme and a licence.

The prevailing way to build a computer-use agent pairs a model with a harness made for one domain: a browser or a desktop environment, fitted with tools that humans engineered and fixed before the task begins. As models get better at writing code, that arrangement starts to hold them back. A harness that is native to the GUI and static stops the model from operating directly and programmatically on the state of the system, and stops it from building tools flexibly as it goes.

CUAWright is a minimal terminal harness, roughly three thousand lines of code. Bash commands are its sole action interface. A file system is its evolvable space, where the agent creates tools dynamically and manages its context. Instead of a fixed toolkit handed over in advance, the agent gets a programmable surface and builds what it needs. The bet is that a model which is good at writing code is better served by a place to write and run it than by a menu of buttons chosen for it ahead of time.

That is the whole design, and the minimalism is the point. One interface, the shell, covers every action. One space, the file system, holds both the tools the agent writes and the context it manages. Where a GUI harness decides the tools before execution, here the tool set can grow during the task itself. And where a GUI stands between the model and the system, the shell lets the model act on system state directly and programmatically, which is precisely the access the static harness was said to deny it.

Across a wide range of digital tasks, the agent with this minimal programmable interface does substantially better than with GUI or hybrid CLI interfaces. On OSWorld two point zero, CUAWright gives a thirty-three point two percent relative improvement in partial reward, while cutting estimated cost by thirty-seven point five percent, against the published baseline with GPT five point five.

On Online-Mind2Web, it beats the GUI native harness by four point seven percent in success rate. On the long-horizon Odysseys benchmark the margin is forty-four percent. The gains also reach CAD applications, which need accurate visual understanding as well as command-line interaction. On CADGenBench and BenchCAD, the harness yields relative improvements ranging from eight point one percent up to forty-one point six percent over other CLI based harnesses, again with GPT five point five.

The conclusion the authors draw is about the environments rather than the model. Digital environments are far more programmable than their GUI interfaces suggest, and a minimal harness centred on the terminal, they argue, is the key to both better performance and better efficiency. The OSWorld result shows both halves of that at once: more reward for less estimated cost, from a harness that is smaller, not larger.