现在主流做法是截图加鼠标键盘，等于把人的交互方式强加给模型：每一步都要重新读屏，电脑本身的语义层和程序化能力反而没被用上。产品里早就靠 accessibility tree…
This is a AI post classified by AGK Intelligence as AI agents (research), kept by the AI Radar because it carries real work, not commentary.
CUA 的榜单刷得越来越快，但很少有人算时间账。 现在主流做法是截图加鼠标键盘，等于把人的交互方式强加给模型：每一步都要重新读屏，电脑本身的语义层和程序化能力反而没被用上。产品里早就靠 accessibility tree 干活了，研究里 AX 多数时候只被当成 observation，这就是研究和落地之间的 gap。 Hybrid Computer Use 把三个通道统一成一套 interface 交给 general agent，每一步由模型自己选：code 与命令行、AX 语义操作、GUI 视觉操作。工程上有几个关键动作，AX 工具返回紧凑观察、批量执行动作、以 AX diff 报告效果，把往返次数压下去；切换通道时防止拿过期状态去执行；再配一份 skill 教模型怎么组合通道；外层以 MCP 接入任何 harness。 结果在 OSWorld 2.0 上 Partial 和 Binary 两项都是最高，而且轨迹分析显示所有被测模型都受益，只是分配在三个通道上的工作量不同。效率和成功率同时更好，跨平台和跨模型都成立。 我认同作者那句判断：AX 在执行层被严重低估，它是电脑在语义层的表示，和 GUI 某种程度是对偶的，比像素更 LLM native。纯 GUI 是第一代 human bias，code execution 加语义接口更像下一代。 也要说边界。AX 的质
Posted by 青稞社区 (3.7k followers) 14 h ago · 6 likes · 486 views · view the original post on X. Kept by the AI Radar as AI agents. Tools mentioned: arXiv.org.
More AI work like this
- Trying out Gemini 4 Argon inside Antigravity with FreeCAD. You can easily set up the MCP… — @ai_for_success
- Your agents can now respond to real-world events with Ring Signals. — @mastra
- this week at langchain: — @hwchase17
- Day 18/ — @hanghuang_
- Get @bot - it will change your life. Then add the open Longevity Grok and it will work… — @biogerontology
- If you want to learn how to build self-improving harnesses, check out this paper. — @dair_ai
- AI agents are eating good this week 🤑 — @coingecko
- I asked what the oldest or weirdest device talking to your Hermes Agent is. my favorites: — @witcheer
Every post is read and classified by AGK Intelligence: what it is, which market it belongs to, and whether the link is a real tool. 65.3k posts from 5k X accounts over the last 14 days, 10.4k tools, 19 markets. New posts are collected several times an hour and the full ranking is recomputed every hour — last update 2026-11-06 17:49 UTC. Full method.