Cua

Overview

Cua provides infrastructure for computer-use agents. Its product overview describes fleets of machines for training, evaluation, and data generation across Linux, Windows, macOS, and Android. The platform emphasizes running many agent sessions, restoring snapshots, and reproducing failures.

The published product family includes Cua Driver, Sandbox, Fleets, Bench, and an environment catalog. Driver concerns interaction with desktop applications; Sandbox provides isolated computer environments; Fleets coordinates larger pools of machines. The site also identifies open-source control and evaluation components.

Cua's role in the RL supply chain is primarily execution infrastructure. It gives agents access to computers and provides machinery for operating those computers at scale. Whether a specific engagement also includes authored tasks, verified demonstrations, or evaluation content should be established separately rather than inferred from the availability of a runtime.

Why computer infrastructure is different

Our assessment is that graphical-agent workloads require more than the ability to execute code. An agent may need to observe a window, click a control, wait for an application to respond, and verify that a file was saved. Timing, screen state, and operating-system behavior become part of the experiment.

An illustrative cross-platform evaluation might ask the same agent to complete comparable file-management tasks on different desktops. Differences in outcomes could reflect the model's visual understanding, the interaction driver, or the application itself. Keeping those layers observable helps a team diagnose the actual bottleneck.

Evaluating fit

Start with the operating system and applications required by the intended workload. Test full tasks rather than only machine startup or a single successful click. Measure action latency, session stability, reset behavior, and the ability to recover logs and screenshots from failures. For large batches, examine how queued jobs and partially completed runs are handled. Snapshot support should be tested against the application state that matters to your task. Buyers comparing hosted and self-managed options should include maintenance and debugging effort in the calculation. A working computer fleet is valuable infrastructure, but task quality and outcome verification still determine the usefulness of the resulting training experience.