Undergrad @ BUPT, Intern @MSFTResearch Asia General Agent, Computer Use Agent, LLM
Beijing
Joined December 2025
- Tweets174
- Following1K
- Followers121
- Likes1.4K
Strongly agree. RUA is retracing CUA’s path: SoM and grounding models gave way to base models. Only base models have long-horizon ability and world knowledge, so RUA ends up there too. Latency won’t block it: faster inference, System 1 actions, shorter reasoning (like Astra).
Claim: Computer-Use agents are embodied generalists
Back in August, we used a single computer-use agent across 4 different embodied AI-related competitions at ECCV, and we ranked 1st, 2nd, 3rd, and 6th places.
How? We constructed simple GUI software for each competition that connects to the benchmark tasks (e.g., using a keyboard to navigate inside a 3D house rendered in a web interface), and that's it. The agent itself stays the same.
I have been meaning to write about it for a long time, but here it is now:
pranjal2041.github.io/cua-em…
Now to the position: we believe that computer-use agents are embodied generalists. Why?
TLDR: We believe computer-use agents share striking similarities with models for embodied AI, and embodied tasks can be formulated as interactions with digital software. Once we have that, we can learn from what worked for computer use over the past 2 years, focus on digital scaling, and avoid mistakes or local minima in the grand scheme of things.
I would urge anyone working in computer use or embodied AI to take a quick read of the paper, or at least go through the website. We do not think many people would agree now, and only time will tell.
But we do think this is an important line of thought that everyone should consider. We look forward to feedback and constructive criticism!
Haoqing Wang retweeted
Claim: Computer-Use agents are embodied generalists
Back in August, we used a single computer-use agent across 4 different embodied AI-related competitions at ECCV, and we ranked 1st, 2nd, 3rd, and 6th places.
How? We constructed simple GUI software for each competition that connects to the benchmark tasks (e.g., using a keyboard to navigate inside a 3D house rendered in a web interface), and that's it. The agent itself stays the same.
I have been meaning to write about it for a long time, but here it is now:
pranjal2041.github.io/cua-em…
Now to the position: we believe that computer-use agents are embodied generalists. Why?
TLDR: We believe computer-use agents share striking similarities with models for embodied AI, and embodied tasks can be formulated as interactions with digital software. Once we have that, we can learn from what worked for computer use over the past 2 years, focus on digital scaling, and avoid mistakes or local minima in the grand scheme of things.
I would urge anyone working in computer use or embodied AI to take a quick read of the paper, or at least go through the website. We do not think many people would agree now, and only time will tell.
But we do think this is an important line of thought that everyone should consider. We look forward to feedback and constructive criticism!
Haoqing Wang retweeted
1/ Today we're launching Cua Driver Bench, together with Scale AI's AgentEnv.
Can an agent do a desktop task in the background while you keep working undisturbed? On macOS, Cua Driver fully passed 58%, Codex computer use 33%, Vercel's agent-desktop 20%, browser-use 3%.
Proud to be a contributor to WebHarbor! ⚓️ Congrats to Zhaoyang and the whole team 🎉
It laid the groundwork for Mage-VL-Cowork: replica sites for synthesizing web tasks, resettable environments, and reliable offline evaluation that doesn't depend on live websites. Great work!
⚓ WebHarbor⚓ Introducing the largest open-source environment resources for GUI Web Agents!
🌐 153 real-world websites (Amazon, GitHub, Booking, YouTube, ...)
🎯 3,518 tasks, including complex multi-step and authenticated workflows
🤖 Built for reproducible evaluation and scalable RL training
We started with just 15 environments in May 2026. Today, thanks to 57 contributors from 26 institutions, we've scaled to 153 high-fidelity, image-rich environments—the largest open-source collection of its kind, to our knowledge! 🔥
And they're all executable, offline, and fully open-source!
👀 Explore the environments now: aiming-lab.github.io/WebHarb…
Why did we build it, and how? 🧵👇 Threads 1/5
Thanks Senqiao for sharing, and for building the Mage-VL foundation that made this possible! 🙏
Building on it, we took a 4B checkpoint with no agentic CUA skills to 66.0% on OSWorld-Verified: midtraining, synthetic tasks, harness iteration, SFT, and RL. Full recipe 👇
Frontier models are pushing Computer Use forward. 🚀
But what about latency, privacy, and on-device AI?
Excited to share @Microsoft's open-source Mage-VL-Cowork: a 4B Hybrid CUA built on our Mage-VL pretrained model!
⚡ 66.0% on OSWorld-Verified!
How did we get there? 🧵👇
Haoqing Wang retweeted
Frontier models are pushing Computer Use forward. 🚀
But what about latency, privacy, and on-device AI?
Excited to share @Microsoft's open-source Mage-VL-Cowork: a 4B Hybrid CUA built on our Mage-VL pretrained model!
⚡ 66.0% on OSWorld-Verified!
How did we get there? 🧵👇
Thanks Francesco! Cua Driver's work paved the way for this. We learned a lot from it while designing HCU, both what works and where we could push further. For personal computers and background computer use, Cua Driver is still the go-to choice 👍
Agents can write code, read accessibility trees, and look at pixels. Most computer-use interfaces offer only screenshots and clicks.
Hybrid Computer Use (HCU) gives agents all three channels and lets them pick one per step, pushing performance and efficiency to the frontier.
🧵
6/ HCU is open source (MIT) as an MCP server, SDK, and CLI. It plugs into any harness: Codex, Claude Code, Gemini CLI, Grok Build, and more, on Linux, macOS, and Windows.
Paper: arxiv.org/abs/2610.41432 Code: github.com/microsoft/Hybrid-…
7/ Thanks to my amazing collaborators Xiaoyi, Sicheng, and Zhexin!
An upvote on HF Papers or a star on GitHub would mean a lot 🙏 huggingface.co/papers/2610.4…
Haoqing Wang retweeted
Microsoft just released a 4B hybrid computer-use agent on Hugging Face
One compact policy that knows when to click — and when to code
66.0% on OSWorld-Verified, ahead of 27B and 30B agents
Thanks a lot for the demo, Niels! 🙏 So cool to see it running on HF.
Microsoft released a new 4B model called Mage-VL-Cowork that sets a new SOTA for on-device computer use 🔥
I built a demo that lets you submit a task like "Go to Wikipedia and find out when the Eiffel Tower was completed." Watch it click, type, scroll, and use a terminal on a real Linux desktop.
The whole thing runs on @huggingface: the model in a Space, the desktop in an HF Sandbox. No VMs to manage. 🧵
0:39 vs 9:56.
Same spreadsheet task. Our 4B model drops into the terminal and writes Python in 6 steps; a 27B GUI-only agent clicks through 39.
Meet Mage-VL-Cowork: one model that picks GUI or CLI at every step. 66.0% on OSWorld-Verified. Open weights.
🧵 0/7
6/ Everything to reproduce the 66.0% is open: Apache-2.0 weights, an SGLang Docker image, and the 361-task eval suite.
Paper: arxiv.org/abs/2610.40747 Page: microsoft.github.io/Mage/cow… Code: github.com/microsoft/Mage HF: huggingface.co/microsoft/Mag…
7/ Joint work with co-first author Xiaoyi Zhang at Microsoft Research Asia.
An upvote or a star means a lot 🙏 HF Papers: huggingface.co/papers/2610.4… GitHub: github.com/microsoft/Mage
This thread is the training side. Tomorrow: the interface side of hybrid computer use.
