COMPUTER USE / OWNER WORKSPACE
Computer Use Lab
Where does a model lose control of an interface? Choose canvas, browser or album-editing tasks and save states and events as it works. The lab helps examine specific failures, from dragging a narrow handle to waiting for an asynchronous interface.
For people evaluating computer-use agents and comparing their behavior or settings on the same tasks.
The workspace and guides require owner access through Cloudflare Access. Saved runs and snapshots are private.
Choose tasks, record interaction, inspect the outcome
- Select one task, a subset or the full suite. For a comparison, hold the task set, version, seed, viewport and interaction language constant.
- Let your model operate the task interface. Review the saved events and snapshots to trace where the state changed.
- After all selected tasks finish, enter the model name, reasoning setting and reported token count, then finalize. The application evaluator grades the tasks together.
Model details and token counts are reported by the operator. These focused tasks do not establish performance across real work. External grading requires a configured connection; missing configuration or failed calls do not count as completed external grading.
Choose an evaluation environment for the behavior you need
| Environment / edition | What it evaluates and conditions to check |
|---|---|
| Computer Use Lab | Focused GUI interactions within this lab. Check the shared tasks, initial state, version, seed and viewport alongside interaction records and evaluator outcomes. |
| OSWorld v1 / OSWorld- | Real OS tasks involving web and desktop apps, including work across apps. Verified refines the v1 tasks and evaluation infrastructure. Identify the release, setup and evaluation conditions used. |
| WebArena (original) | Tasks on self-hostable websites, evaluated for functional correctness. Check site state, observation format and task-specific evaluation conditions. |
| OSWorld 2.0 | The current successor targets long workflows and dynamic environments. Treat its tasks and evaluation as distinct from v1 / Verified. |
Official sources checked: . Original papers: OSWorld · WebArena · OSWorld 2.0