Training an agent to click, type, and navigate like a person currently means wiring together pieces that were never designed to fit: one repo for the agent, another for the environment, a third for the traces, and a fourth for evaluation. A UC Berkeley team argues that fragmentation is why computer-use research moves slowly, and its answer is CUA-Lite, an open platform that standardizes the whole loop behind one action space, one data schema, and one command across desktop, browser, and mobile.
The project’s headline piece is Lite.OSWorld. The familiar OSWorld benchmark ships each task as a full QEMU virtual machine, which demands nested virtualization that most cloud fleets never expose. CUA-Lite rebuilt the same suite and the same graders to run on a GNOME desktop inside a plain Docker container with no /dev/kvm requirement, and reports matching scores across 13 models, meaning a training signal earned in the container still transfers to the official benchmark.
Related sandboxes now ride the same base: Lite.ScaleCUA, Lite.CUAGym, and Lite.CUAWorld, the last reaching roughly 40 applications from Blender and QGIS to VS Code, for more than 30,000 verifiable tasks in total. A layer called LiteSample defines one supervised-learning schema stored as parquet files plus images, and more than ten public datasets, among them Aguvis, OpenCUA, ScaleCUA, and Multimodal-Mind2Web, have been converted and released free on Hugging Face, alongside fresh rollouts generated by a frontier teacher model for distilling into smaller student agents.
Environments connect through lite.gym, where screenshots travel up and actions travel down, and per-model adapters translate the shared schema into each family’s native training format. Setup runs through uv sync on Python 3.12, so a full research loop now fits on any Docker host, including CI runners and nested containers.
