CUAWright Runs Computer-Use Agents Through a Single Shell Tool

Researchers led by Yadong Lu, with collaborators from The Ohio State University and Microsoft Research, have posted a preprint describing CUAWright, a computer-use agent harness built around one action: run_command. The model uses that command to inspect its environment, write scripts, execute programs, manipulate files, and invoke keyboard or mouse controls when direct automation is unavailable.

Across five browser, desktop, long-horizon, and CAD benchmarks, the authors report higher task scores and lower estimated costs than GUI-native and hybrid command-line baselines. The design moves tool creation into the agent’s runtime, allowing a coding model to build task-specific utilities instead of relying on a fixed collection of browser and desktop actions.

Tooling moves into the task

Computer-use agents usually depend on a harness, the software layer that translates model output into clicks, keystrokes, browser actions, and observations. Many harnesses repeatedly capture screenshots, append them to the model’s context, and expose low-level primitives such as click, type, scroll, and drag.

A screenshot-driven loop reveals only the visible interface and can consume substantial context as a task grows. Low-level actions also require many model turns to complete work that a script could handle in one operation, such as inspecting page elements, transforming a spreadsheet, or editing a structured file.

CUAWright’s harness contains roughly 3,000 lines of code and gives the model a shell plus a scratch filesystem. The model can create new tools while solving a task, save intermediate state outside its limited context window, and retrieve specific information later with ordinary command-line utilities.

One command drives the loop