Does a computer-use agent need to do everything with a mouse? Rearranging a spreadsheet through dozens of clicks can be a detour when a short program could make the change directly. Inspecting a drawing or interacting with an application's live state may still call for the screen. Microsoft's released Mage-VL-Cowork is trained to choose between those routes.
The official model card describes a vision-language policy presented as a 4-billion-parameter model, combining graphical actions with a persistent terminal. Weights are available under Apache 2.0. There is a catalog discrepancy: Hugging Face's automatic counter displays 5B, while the authors' documentation uses 4B. This article follows the team's technical description.
Choosing a tool becomes part of the task
The official repository documents two channels: computer_use for screenshots, mouse and keyboard, and bash for shell commands and Python. One policy chooses the channel without an external model router.
Imagine preparing a report from a spreadsheet: transform the data with code, then open the result and inspect its presentation. That illustrates the possible value of hybrid execution; it is not an outcome measured by this publication. Taking a more direct route still leaves the agent responsible for checking the result.
That changes what matters in a demonstration. Counting clicks or commands on their own says little. A short sequence can modify the wrong file, while a longer one can finish correctly. A useful evaluation connects execution to the finished artifact and observes how the agent recovers when something goes wrong.
What the reported 66% tells us
The team reports 66.0% on OSWorld-Verified across 361 tasks, excluding Google Drive, with a 50-step budget. The figure averages two full runs. This is the authors' evaluation under their documented conditions; MnzAI Labs has not independently reproduced it.
The result provides a starting point for experiments. It does not mean the model will complete two thirds of anyone's everyday tasks, or establish that it can safely operate real accounts. Different applications, files and permissions can change the difficulty substantially.
The documentation calls for isolated, disposable environments and human confirmation before irreversible actions. That fits the reach of a shell: the channel that simplifies file editing also increases the consequences of an incorrect instruction.
An evaluation can begin with document copies, a virtual machine and success criteria established in advance. Measuring total time, errors, interventions and file integrity helps answer the useful question raised by this approach: does switching between screen and code improve the delivered work, or simply change what automation looks like?