ComputerRL

Overview

ComputerRL is a framework for scaling online reinforcement learning for computer-use agents across distributed virtual desktops. It combines API and GUI interaction and focuses on training stability at scale rather than only evaluation.

Why it matters

It matters because a harness does not really have a gym until it can run many environments in parallel without turning infrastructure into tragic performance art.

Distinctive trait

Its distinctive trait is infrastructure seriousness: thousands of parallel desktops, mixed API/GUI interaction, and training strategies to survive long online runs.

Relationships

Read ComputerRL with osworld, windows-agent-arena, appworld, and the training-focused section of rl-gyms-and-executable-environments-for-ai-harnesses.