simular-ai/Agent-S ? reverse-engineered prompt

Build me an open source computer use agent that can take a plain English task and then use the mouse, keyboard, and screen like a person to get it done in normal desktop and web apps. I want it to work across macOS, Windows, and Linux, and to feel like a real framework I can extend, not just a demo.

Please make the app able to look at screenshots, decide the next action, and keep going until the task is finished, with a simple way to plug in different model providers. If it makes sense, include basic memory, planning, and retrieval so it can handle longer tasks more reliably. Also add clear setup instructions, a few example tasks, and tests or evaluation helpers so I can see it working. If you need current implementation details from the docs or related packages, look them up online and follow the latest guidance.