Jev-windows-agent
A Windows UI Automation backend for computer-use agents that clicks, types and navigates real Windows applications through a fast, structured decision loop instead of a screenshot-and-guess approach, so a frontier model is not needed for every click.
jev-windows-agent is a Windows UI Automation extension of an existing computer-use agent project, arc-cua, adding a perception and execution backend for Windows applications such as Notepad, Settings, Spotify and File Explorer. Rather than a large-language model reasoning about every individual click through a repeated observe-reason-act loop, a planner hands off a bounded subtask, described as a small payload that any GPT, Claude, Gemini or local model can generate, to a fast decision model that executes the whole UI sequence and returns control once the subtask completes.
It keeps the original project's runtime, validation and decision policy unchanged while adding the Windows-specific backend, transport support for routing model calls through OpenRouter, and an optional confidence-gating step. The stated optimization target is fewer expensive reasoning calls per completed task, not fewer individual UI actions.