Richard Lee

[Ang Li @angli_ai]

I left DeepMind to give AI a body.

The next billion computers will be autonomous, driven by agents.「box」by @SimularAI is our step toward that: a cloud computer every agent can share.

「box」is one piece of the infrastructure in our autonomous computer stack: cloud computer + computer use capabilities. It already generated great interest pre-launch, and we are excited to share it with the public today.

Try it for free!

[Simular @SimularAI]

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.

[Richard Lee @Richard_simular]

@sama hey sam your support team left me + someone from oai on read for 3 months on a ticket
Its weird now I have 20 emails between him and i and not hearing back from support ;/

[Yuzhe Yang @Toby_Yang_7]

Check out our latest work on computer-use agents, Agent S4! We believe the future of CUA is not just stronger models, but agents that build and evolve their own interfaces. By learning from experience and improving how they interact with computers, agents can become more capable and efficient. Excited to share this fully open-source work!

Huge thanks to @saa1605 , @chalo2000, @Bhuvan_Koduru, @Richard_simular, JC, @angli_ai, @xwang_lk, and the @SimularAI team for making this happen! It’s been a pleasure working together and bringing this idea to life.

[Xin Eric Wang @xwang_lk]

People long believed in human-like computer use. But software (even terminals) is built for humans, not AI agents. Why not let agents build their own?

Meet 𝗔𝗴𝗲𝗻𝘁 𝗦𝟰, an agent scientist evolving its own interfaces, letting an open 27B model beats GPT-5.5 & Opus 4.8.

With its evolved interfaces, the same model improves 𝟯.𝟴× and costs <𝟱% as much per task as either frontier model.

Agents today read screenshots and act through click, scroll, type, maybe a shell. So an agent fills a spreadsheet cell by cell when one line of code would do, or prints two thousand lines of a file to find one function.

Every new app or model release means a human reads trajectories, diagnoses failures, and hand-writes new primitives. Skill-distillation methods can reorder what the interface allows, but they can’t extend it. Agents compensate with more test-time compute instead of a better interface.

𝗠𝗼𝗱𝗲𝗹𝘀 𝗶𝗺𝗽𝗿𝗼𝘃𝗲; 𝗶𝗻𝘁𝗲𝗿𝗳𝗮𝗰𝗲𝘀 𝘀𝘁𝗮𝘆 𝗳𝗿𝗼𝘇𝗲𝗻.

“We want AI agents that can discover like we can, not which contain what we have discovered.”
— Rich Sutton, The Bitter Lesson

Hand-designed interfaces are what we have discovered. Agents can read their own trajectories, spot what’s missing, and write and test code. 𝗜𝗻𝘁𝗲𝗿𝗳𝗮𝗰𝗲 𝗱𝗲𝘀𝗶𝗴𝗻 𝘀𝗵𝗼𝘂𝗹𝗱 𝗯𝗲 𝗮𝗻 𝗮𝗴𝗲𝗻𝘁 𝗰𝗮𝗽𝗮𝗯𝗶𝗹𝗶𝘁𝘆.

That’s what Agent S4 does.

Agent S4 makes a 𝗿𝗲𝗽𝗼𝘀𝗶𝘁𝗼𝗿𝘆 𝗼𝗳 𝗶𝗻𝘁𝗲𝗿𝗳𝗮𝗰𝗲𝘀 𝗮𝗻𝗱 𝘂𝘀𝗮𝗴𝗲 𝗸𝗻𝗼𝘄𝗹𝗲𝗱𝗴𝗲 the unit of evolution. A task agent attempts synthetic tasks with no success labels or verifiers. A meta-agent studies the trajectories, forms hypotheses about which actions, observations, or knowledge would help, and builds descendants through mutation or crossover. Validation tasks the meta-agent never sees decide which repositories survive.

The model and harness stay fixed, and the meta-agent 𝗻𝗲𝘃𝗲𝗿 𝘀𝗲𝗲𝘀 𝗮 𝗯𝗲𝗻𝗰𝗵𝗺𝗮𝗿𝗸 𝘁𝗮𝘀𝗸. 𝗢𝗻𝗹𝘆 𝘁𝗵𝗲 𝗶𝗻𝘁𝗲𝗿𝗳𝗮𝗰𝗲 𝗲𝘃𝗼𝗹𝘃𝗲𝘀.

With the evolved repository, Qwen3.8-27B goes from 𝟱.𝟲% to 𝟮𝟭.𝟯% on OSWorld V2, 𝟯.𝟴× the same model with fixed interfaces and a new state of the art for open-source computer-use agents.

It beats GPT-5.5 (13.0%, $25.50/task) and edges past Opus 4.8 (20.6%, $72.40/task) at $𝟭.𝟭𝟬 per task.

On ScienceBoard, the same model goes from 41.4% to 58.6%.

The analysis is the more interesting part. Evolution mostly deepens interfaces the agent already has: 4 new interfaces and 127 extensions. Give the meta-agent a scratch VM to test its code, and it builds 12 new interfaces and 253 extensions, reaching music scores, audio projects, and 3D scenes. The best repository came from crossing two earlier branches.

In one case study, the seed agent hand-builds a browser client mid-task and debugs through failure after failure; the evolved agent calls page_text and reaches the same message at tool call 𝟭𝟭 instead of 𝟯𝟳.

Human-like computer use was a starting point, not the destination. The interface an agent acts through is a second axis of improvement, beyond better models and harnesses, and agents can now design it themselves. 𝗘𝘃𝗼𝗹𝘃𝗶𝗻𝗴 𝗮𝗴𝗲𝗻𝘁 𝗶𝗻𝘁𝗲𝗿𝗳𝗮𝗰𝗲𝘀 𝗶𝘀 𝗮 𝗻𝗲𝘄 𝗱𝗶𝗺𝗲𝗻𝘀𝗶𝗼𝗻 𝗼𝗳 𝗮𝗴𝗲𝗻𝘁 𝘀𝗲𝗹𝗳-𝗶𝗺𝗽𝗿𝗼𝘃𝗲𝗺𝗲𝗻𝘁.

[Simular @SimularAI]

Enterprises and individuals increasingly find computer-use agents powerful. But the cost of scaling usage weighs on them.

That's because these agents call on LLMs to reason every step and repeat. Our CTO Jiachen Yang told @miyadavid at @ISMG_News how we use neuro-symbolic to bring costs down:

The agent learns a workflow once, then repeats it with code, so it doesn't run a frontier model on every step.

On OSWorld, our @sai_borg was 217x cheaper than Opus alone, with a 72% success rate. Savings start after the third run, and Sai checks before each action.

As Jiachen points out, a single LinkedIn recruiting task costs at least 50 cents with today's best models. "If a recruiter wants to do it for thousands of people, I'm pretty sure at some point it will be unjustifiable in terms of the cost."

The full story here 👇https://t.co/nqybn7ryBo

[Simular @SimularAI]

Ever wondered what goes into making @sai_borg more reliable?

@Richard_simular from @SimularAI’s research team was invited to BayLearn 2026. His presentation explored how computer agents can reuse successful steps and handle repetitive tasks more consistently.

For more on how agents learn reusable workflows, read the paper Richard coauthored: https://t.co/DwU3yqOVMY

Thanks to Santa Clara University for hosting, and to San José State University, UC Berkeley, and UC Santa Cruz for co-organizing the event 👏

#BayLearn2026 #AIResearch

[Xin Eric Wang @xwang_lk]

𝐁𝐞𝐭𝐭𝐞𝐫 𝐫𝐞𝐬𝐮𝐥𝐭𝐬 & 𝐥𝐨𝐰𝐞𝐫 𝐜𝐨𝐬𝐭, 𝐚𝐬 𝐬𝐨𝐨𝐧 𝐚𝐬 𝐲𝐨𝐮 𝐩𝐮𝐭 𝐟𝐫𝐨𝐧𝐭𝐢𝐞𝐫 𝐚𝐠𝐞𝐧𝐭𝐬 𝐨𝐧 𝐋𝐄𝐀𝐒𝐇! 🦮

Why use a frontier model at every step when a lightweight agent can handle lots of the work?

Introducing 𝐋𝐄𝐀𝐒𝐇: 𝐔𝐧𝐜𝐞𝐫𝐭𝐚𝐢𝐧𝐭𝐲-𝐀𝐰𝐚𝐫𝐞 𝐖𝐞𝐚𝐤-𝐭𝐨-𝐒𝐭𝐫𝐨𝐧𝐠 𝐄𝐬𝐜𝐚𝐥𝐚𝐭𝐢𝐨𝐧 𝐚𝐧𝐝 𝐎𝐯𝐞𝐫𝐬𝐢𝐠𝐡𝐭.

LEASH unleashes the strong agent only when the weak agent's latent uncertainty signals the need, and keeps it on a leash afterward.

Two key challenges:
1️⃣ 𝐖𝐡𝐞𝐧 𝐭𝐨 𝐞𝐬𝐜𝐚𝐥𝐚𝐭𝐞? Use the weak agent's latent uncertainty as an inexpensive signal that can be monitored at every step.

2️⃣ 𝐇𝐨𝐰 𝐭𝐨 𝐦𝐚𝐢𝐧𝐭𝐚𝐢𝐧 𝐨𝐯𝐞𝐫𝐬𝐢𝐠𝐡𝐭 𝐚𝐟𝐭𝐞𝐫 𝐞𝐬𝐜𝐚𝐥𝐚𝐭𝐢𝐨𝐧? Even frontier agents can go astray. And when served through black-box APIs, their internal states are inaccessible. LEASH enables continued oversight without accessing the strong agent's internals.

𝐓𝐡𝐞 𝐤𝐞𝐲 𝐢𝐧𝐬𝐢𝐠𝐡𝐭: Knowing when to escalate—and when to intervene—is just as important as knowing how to act.

The future of agentic AI isn't about using the strongest model everywhere. 𝐈𝐭'𝐬 𝐚𝐛𝐨𝐮𝐭 𝐝𝐞𝐩𝐥𝐨𝐲𝐢𝐧𝐠 𝐢𝐧𝐭𝐞𝐥𝐥𝐢𝐠𝐞𝐧𝐜𝐞 𝐰𝐡𝐞𝐫𝐞 𝐢𝐭 𝐦𝐚𝐭𝐭𝐞𝐫𝐬 𝐦𝐨𝐬𝐭.

[Richard Lee @Richard_simular]

@pavliscak @SimularAI @sai_borg Definitely!

[Ang Li @angli_ai]

@Richard_simular presenting neuro-symbolic computer use agents at @BayLearnOrg today!

[Richard Lee @Richard_simular]

oops

[Ang Li @angli_ai]

When a teammate gave up talking to me about reviewing their PRs ..

[Ang Li @angli_ai]

Build your own personal agent within 10 minutes, using the Autonomous Computer API by @SimularAI

It also self optimizes token usage for repetitive tasks btw

https://t.co/OjssM8SI8h

[Simular @SimularAI]

Our MTS @Richard_simular, part of the team that built @sai_borg, answered 4 critical questions on personal agents on a panel at #SFTechWeek:

1. When should an agent think, when should it just act, and when must it stop and ask — what’s your decision architecture?

Start super strict then slowly relax the guard. User & personal agent have a continual "bonding" process, and people have different preferences so it's never going to a one size fits all

2. Do agents overthink into paralysis?

Usually it's not an agent, but a weak model overthinking.
Some models do this to benchmaxx so you will see an agent thinks for 4 minutes for a simple task like open https://t.co/J4hgwhuLn1.

3. How do we know how much autonomy to give to an agent?

Running on sandbox is (almost) always the best approach. Factors: 1. how safe is sandbox 2. how reachable is the request to you 3. how quickly it learns and relax permission

4. With Muse Gadgets and OpenAI Dots launching last week, what does personal-agent decision architecture look like 12 months from now

How to give more information to agents. If we can find out a way to send explicit and exact request to agent then we don't need any context.

Thanks for having us on!

[Richard Lee @Richard_simular]

Insights on building reliable agents (they also happen to be very cost effective!)

[Ang Li @angli_ai]

Three years into making computer use agents, I kept asking myself:

Why hasn't CUA reached super intelligence yet?

@sai_borg (Sai), which happens to sound like SI, has invented a new way to build computer use agents: not just more reliable, but also 99% cheaper

The answer lies in how humans function: build muscle memory to optimize for efficient output over time

We released a paper today on this new paradigm, called neuro-symbolic

Give it a read here: https://t.co/Omjqm0t11W

[Richard Lee @Richard_simular]

@miltonheyan @Techweek_ @CoreSpeedHQ @SimularAI @ZongZiWang @harvey @haosu_cs @samuel_spitz @Replit Thanks for having me, excited to see you guys next week!

[Ang Li @angli_ai]

Our @Richard_simular will speak here and tell why your agents shouldn’t overthink too much

[Richard Lee @Richard_simular]

@ZeningChen42844 Can it help me to climb grandmaster in league 👀

[Richard Lee @Richard_simular]

@sai_borg @sai_borg can i get 5 more vids today
https://t.co/DWODratCij

[Richard Lee @Richard_simular]

Day 1 of assembling my gtm team with Sai agent 👀
#SaiFleet #Robosecretary

[Richard Lee @Richard_simular]

Day 1 of assembling my gtm team with Sai agent 👀
#SaiFleet #Robosecretary

[Xin Eric Wang @xwang_lk]

There are plenty of “human-pretending-to-be-an-agent” game demos out there.

This one is real.

Sai + Jev played Minecraft, fought mobs, and spawned mining sub-agents—all in real time.

1 hour. $0.01.

And Sai can spin up 100× agents in parallel for under $1 total.

#RoboSecretary #SaiFleet