Mage-VL-Cowork: A Training Recipe for Hybrid Computer-Use Agents

Introduction

Computer-use agents are increasingly expected to do more than recognize a screen and click a button. A useful agent may need to inspect a graphical interface, write a short program, invoke a terminal command, and then return to the GUI while preserving the task state. Mage-VL-Cowork addresses this mixed setting with a 4B hybrid computer-use agent and a training recipe that spans data generation, supervised learning, and reinforcement learning.

The work starts from a generally multimodally aligned checkpoint rather than presenting computer use as an isolated capability. Its central premise is that visual interaction and programmatic execution should be learned as parts of one policy. The paper describes a progressive five-stage curriculum covering midtraining, SFT, and RL, although the supplied material does not include the full stage-by-stage configuration.

Key points

• A progressive curriculum for mixed actions. The training process is designed to combine GUI operations, code, and terminal execution instead of training each interface mode as a separate skill. This is important for tasks where the most efficient action can change during execution.
• A reusable data-production system. Mage-CUA-Forge replaces repeated, task-specific agent construction with skill-conditioned generation workflows and application-specific preparers. The system reportedly synthesizes 293,806 desktop and web tasks across 6,915 skills while reducing recurring per-task expenditure by 82.6%.
• Harness evolution as part of training. Rather than directly copying trajectories from an unchanged teacher, the authors evolve the teacher harness and instantiate specialized CUA harnesses for different applications. These harnesses are intended to produce stronger trajectories and supervision tailored to different training stages.
• A compact benchmark result. Mage-VL-Cowork-4B reaches 66.0% on OSWorld-Verified. The abstract characterizes this as state-of-the-art at its parameter size and also reports more efficient reasoning cost.

Why it matters

The most notable aspect of the work is its emphasis on the training system around the model. Computer-use agents require broad coverage of applications, skills, and tool combinations. Producing that coverage manually can become a bottleneck, while generic demonstrations may fail to teach when an agent should switch from visual interaction to code or terminal execution. By introducing reusable generation workflows and application-specific preparation, Mage-CUA-Forge targets the cost and consistency of large-scale task creation.

The harness idea is also significant. In this setting, the teacher is not merely a model that emits answers; its environment, tools, recovery procedures, and application adapters affect the quality of the trajectories it can generate. Evolving that surrounding harness may therefore provide a more direct way to improve supervision than simply collecting more outputs from a fixed setup.

The result points toward a route for capable yet relatively small computer-use models, including possible on-device deployment. Still, the supplied source is limited to the paper abstract. It does not provide ablations, task-level distributions, failure analysis, or independent reproduction details. The 66.0% score and the paper’s claim of leadership at its size should therefore be interpreted alongside the full paper and later evaluations. The broader lesson is that progress in computer use may depend as much on scalable data factories and training environments as on model size alone.

arXiv