Agent stealing your mouse? Give it its own computer
Codex uses Sai MCP to operate a calculator on its own on a hosted Windows desktop, no longer taking the developer's mouse. Simular's Agent S4 is even more counterintuitive: according to their report, without changing the model weights and only improving the interface, an open-source 27B model surpassed GPT-5.5 and Claude Opus 4.8 on OSWorld V2.
Computer-using agents have a problem: they share the same computer with you. You use Chrome, and so do they. While you're writing code, they suddenly grab the mouse. Mouse-grabbing sounds like a joke, but once a developer's attention is broken, their train of thought is completely gone.
A developer tried an approach on X: give the agent its own computer. He connected Codex to Simular's Sai MCP, gave it a task in plain English, and then watched Sai open the calculator on a Windows desktop running in the cloud, enter the calculations one by one, with both results left in History.
MCP is the Model Context Protocol, an interface standard that lets AI models connect to external tools and data. Sai packages computer and desktop control as an MCP service, so developers can plug it into an agent and use it without building and maintaining their own virtual machines, screenshot pipelines, and action loops. One developer pointed out that the real value of desktop access is that agents can handle long-tail tasks in existing applications, and every team doesn't have to reinvent its own computer-control wheel.
Sai is a Simular product. Simular also has a research effort more worth paying attention to, Agent S4, and its approach is a bit counterintuitive: keep the model and agent framework fixed, and only change the interface the agent uses to work. Simular said in a tweet that with this method, the open-source 27B model Qwen3.8-27B completed 21.3% of the 108 tasks in OSWorld V2, surpassing GPT-5.5 and Claude Opus 4.8, with the model weights completely untouched. The capabilities didn't change; the interface became more convenient, and the results were different.
Simular also released another set of numbers: Sai paired with Claude Opus 5 scored a maximum partial score of 79.38% on OSWorld 2.1, at a cost of $14.34 per benchmark task. OSWorld 2.1 is a benchmark for evaluating computer-use agents. These are vendor-reported numbers, not independently verified.
That developer said the biggest benefit of giving an agent an independent execution environment is that your cursor is no longer taken away. He joked that this is especially friendly to people with ADHD, because you don't have to sit and wait while the agent operates. One commenter replied more directly: give it a separate monitor, and each does its own thing. Two hands on one mouse never ends well.
Giving an agent a computer is only the first step. The interface is the bigger variable.
发布时间: 2026-10-26 00:41