San Francisco Bay Area
Joined February 2022
- Tweets889
- Following6
- Followers9.6K
- Likes892
Pinned Tweet
2,000,000,000 PCs sit on people's desks today.
The next billion computers?
They’ll be autonomous computers, driven by agents.
Introducing box by @Simular: one computer for ChatGPT, Claude, Codex, Gemini, DeepSeek and all your other agents.
With box, agents open apps, run the terminal, fill in forms, and stay signed in for whatever tasks come next.
Great question! Your logins live in the browser on your own box, not with any of the AIs.
When you sign in from the box iPhone app, your password goes straight to the box and is never shown to the AI.
An AI can only use your box after you approve it on a sign-in page, where you pick which boxes and for how long.
Every connected AI is listed at box.simular.ai/access and you can revoke any of them in one click. Only connect AIs you trust, since they can use the sites you're signed in to on that box.
Have fun with box!
Every AI you use can now share one box.
We told Claude “you’re white” and Codex “you’re black,” and they played chess on the same Box.
This also means you can pass a task to Claude if your Codex’s hit limit. Token saving anyone?
3
32
404
box is brought to you by Simular, founded by ex-DeepMind researchers.
Our agent reached #1 on OSWorld at 80.6%.
box is free to try. Paste one line into your agent:
“Set up box from box.simular.ai/install”
#simularbox #boxbysimular
box.simular.ai/?utm_source=x
Thanks @a16z for the shoutout to box!
Give your agents computer use, and a box 🥡
Companies spent years battling bots. Now they’re fighting for their business.
AI agents are the majority of Web traffic and generate most code pushed to production. Soon, they'll be buying most of what gets sold online.
New piece with @omooretweets on the B2A economy 👇
Simular retweeted
🪄 If your computer cursor is moving on its own, make sure it’s actually getting work done.
Get your halloween Saibot: sai.simular.ai
Try giving your Codex an interface. Having a computer matters.
The interface matters. Give your agent a computer. It stops block me when doing compute use 😂
Connected Codex to Simular’s Sai MCP, sent a plain-English task, and watched Sai operate Windows Calculator on a hosted desktop. It entered the calculations through the app and left both results visible in History.
For developers, having the computer and desktop control available behind MCP is pretty useful. I can hand my agent a desktop task without building and maintaining the VM, screenshot pipeline, and action loop myself.
Simular’s Agent S4 research also caught my attention: keep the model and agent harness fixed, then improve the interfaces the agent works through. Alongside clicks and typing, better tools can turn repetitive actions into a single operation. I think that direction has a lot of potential for vertical agents too.
This video is a small, real Sai product demo; it doesn’t reproduce the S4 experiments. Waiting was cut from the recording.
Separately, Simular reports the highest partial score on OSWorld 2.1—79.38%—for Sai with Claude Opus 5, at $14.34 per benchmark task. Those are vendor benchmark figures, separate from this demo.
1
10
873
Simular retweeted
It’s not revealed in the paper but this is fundamentally learning a world model of the digital world. This paper marks a new era of computer use agents.
Our agent scientist, Agent S4, just got an open 27B model to beat GPT-5.5 and Claude Opus 4.8 on OSWorld V2, with the model's weights left exactly as they were.
With the interfaces S4 evolved, Qwen3.8-27B completes 21.3% of the 108 tasks, the highest binary success of any open-source computer-use agent and 3.8x the same model with fixed interfaces.
It costs $1.10 per task, against $25.50 for GPT-5.5 and $72.40 for Opus 4.8 💸
Full write-up: simular.ai/articles/agent-s4…
Simular retweeted
CUA should back. I once tried Astra's CUA to get me a car insurance quote 🤣. It hit the usage limit before finishing, and burned money at lightning speed. And our Opus 5.5 doesn't seem great at CUA either. There's definitely room here, and a research direction: a model built for long, unstructured prefill input + fast, cheap decode output.
👀 Agent S4 make CUA not really like CUA, it's more like a combined one. Agent will have a low level action (click scroll, and type) but could also come up with better interface to packing some low level action together (and combine with CLI). I think it makes sense and also could potentially benefit all the vertical agent, it could just make the thing works better and cheaper.
(I try CUA in Lightroom with my photo, save my time :)
Our agent scientist, Agent S4, just got an open 27B model to beat GPT-5.5 and Claude Opus 4.8 on OSWorld V2, with the model's weights left exactly as they were.
With the interfaces S4 evolved, Qwen3.8-27B completes 21.3% of the 108 tasks, the highest binary success of any open-source computer-use agent and 3.8x the same model with fixed interfaces.
It costs $1.10 per task, against $25.50 for GPT-5.5 and $72.40 for Opus 4.8 💸
Full write-up: simular.ai/articles/agent-s4…
Give the Sai API a try. It's the autonomous computer API, and every computer-use agent you build gets its own cloud computer.
platform.simular.ai/
The full paper, Evolving Agent Interfaces, is up on arXiv.
arxiv.org/abs/2610.28601
Agent S4 is open source, so you can try it yourself from our GitHub.
github.com/simular-ai/agent-…
Our agent scientist, Agent S4, just got an open 27B model to beat GPT-5.5 and Claude Opus 4.8 on OSWorld V2, with the model's weights left exactly as they were.
With the interfaces S4 evolved, Qwen3.8-27B completes 21.3% of the 108 tasks, the highest binary success of any open-source computer-use agent and 3.8x the same model with fixed interfaces.
It costs $1.10 per task, against $25.50 for GPT-5.5 and $72.40 for Opus 4.8 💸
Full write-up: simular.ai/articles/agent-s4…
