Simular

Simular, a company in Palo Alto with about thirty people and roughly $26.5 million in the bank, thinks this arrangement is backwards. Its pitch, which it prints on its own website with admirable directness, is that it is "The Autonomous Computer Company." The idea is that a computer should be able to operate a computer - open the apps, read the screen, move the mouse, type the keys - the same way a person does, so that the person can go do something else. Ideally something that is not clicking.

This is a category the industry now calls "computer use," and it is having a moment. OpenAI has Operator. Anthropic has Claude Computer-Use. Google has Project Mariner. The interesting thing about Simular is that it was doing this before the moment arrived, and it did the unusual thing of putting its core framework, called Agent S, on GitHub, where it has accumulated close to ten thousand stars. If you want people to trust an AI that can move your mouse, it turns out, a good first move is to let them read the code.

The founders are Ang Li and Jiachen Yang, who met the way a lot of AI founders meet, which is at Google DeepMind, working on the parts of machine learning - reinforcement learning, continual learning - that are about agents figuring out what to do rather than models figuring out what to say. Li has a Ph.D. from Maryland and a resume that reads like a tour of the field: DeepMind, Baidu's autonomous driving unit, Meta AI, Apple. He also, in a detail that will delight anyone who has ever done competitive programming, placed 39th at the ICPC World Finals. The man can click fast. Now he is building software that clicks for you.

What makes Simular worth writing about is not that it makes an agent - everyone makes an agent - but the specific way it worries about the thing everyone should worry about, which is that these agents make things up. An AI that hallucinates in a chat window writes you a wrong sentence. An AI that hallucinates while operating your computer clicks the wrong button, on your actual accounts, with your actual money. The stakes of being confidently wrong go up considerably when the model has a mouse.

Simular's answer is what it calls a neuro-symbolic approach, which is a fancy term for a sensible idea. First, the neural part: let the agent explore, fumble, try things, until it finds a sequence of actions that actually works. Then, the symbolic part: freeze that successful path into plain, deterministic code - a script you can read, audit, and run again tomorrow without wondering whether the model will improvise. The exploration is creative. The execution is boring. Boring, when the thing controls your desktop, is the entire point.

The commercial product built on this is called Sai, described as an "always-on agentic AI coworker." Sai runs on your Mac, your Windows PC, or an isolated cloud virtual desktop, and it does what a diligent colleague would do: uses the apps, navigates the websites, fills in the forms. It can also drop down a level and talk to APIs, run terminal commands, and write code when clicking is not the efficient path. It is, in effect, a worker who happens to be made of software, sitting inside a workspace made of software, doing software.

The use cases Simular talks about are refreshingly unglamorous, which is how you know they might be real. A car dealership uses it to look up VIN numbers. A homeowners' association uses it to pull terms out of PDFs. In 2026 the company demonstrated Sai handling insurance claims processing on Microsoft's Windows 365 for Agents. None of this is going to headline a keynote about the future of intelligence. All of it is the kind of repetitive digital chore that quietly eats a workday. The most honest thing an AI company can do right now is aim at the boring stuff and actually hit it.

And Simular does appear to be hitting things, at least by the measure the field uses to keep itself honest: benchmarks. On OSWorld, a notoriously hard test of desktop tasks, its Agent S2 became state of the art, reportedly beating OpenAI's Operator and Anthropic's Claude Computer-Use. The company reports 72.6% on OSWorld, 90.1% on the WebVoyager browser benchmark, and 71.6% on AndroidWorld for phones. Benchmarks are not products, and a number is not a customer. But in a field where the demos are polished and the reality is often not, a hard benchmark is one of the few places a claim has to survive contact with an adversary.

The money agrees, or at least it is willing to bet. In December 2025 Simular raised a $21.5 million Series A led by Felicis Ventures, on top of an earlier seed round of about $5 million. The investor list is a tell: NVIDIA's venture arm, NVentures, which likes companies that will buy a lot of compute; Basis Set Ventures and South Park Commons, which like technical founders; and Lenny Rachitsky, the product writer, angel-investing in a company whose entire product is using other products. When the person who studies what makes software good puts money into software that uses software, it is worth noting what he saw.

There is a philosophy under all of this, and Simular states it plainly enough that you can quote it back: "We believe humans shouldn't be bound to computers. AI should free us from devices so we can live a better, more human life." You can read that as marketing, and some of it is. But it is also a coherent position about where this technology should point. Most AI right now is designed to keep you at the screen longer - to chat, to browse, to engage. Simular's stated goal is the opposite: to do the screen part so you can leave it. Whether it gets there is an open question. It is at least an interesting one to aim at.