With Doubao bringing a virtual desktop to Windows, its computer-use agent is sidestepping one of the most obvious points of friction: whether users can keep using their own mouse and keyboard while the agent is working. According to the public product description, tasks run inside a relatively separate environment where the agent can read the screen, control the mouse and keyboard, and carry out steps across web pages and multiple pieces of software.
The approach relies on the GUI rather than MCP, APIs, plugins, or a CLI. The upside is that legacy software and websites without an interface can potentially be covered too; the trade-off is just as clear — interface changes, pop-ups, and permission dialogs can all become sources of mis-clicks.
Not Hijacking the Mouse Is Only Step One
A virtual desktop solves parallel use, not controllability on its own. The product says users can watch the operation steps in real time and pause or take over at any moment. How that plays out in practice comes down to three things: whether the task states which app it's about to touch next, whether pausing takes effect instantly, and whether a reviewable log is left behind after a failure.
For people who need to work across web pages, spreadsheets, and cross-software workflows, an isolated environment like this is closer to a usable form than an agent taking over the main desktop. The metrics worth checking next are equally down-to-earth: whether long-running tasks complete reliably, whether users can step in before critical actions, and whether the boundary the isolated environment draws around local files and login sessions is transparent.
Sources: Doubao's public product information, CocoLoop; verified against the Windows virtual desktop, GUI operation, isolated environment, and pause/takeover descriptions.