An AI agent that literally drives your computer for you sounds exciting—but also a little unsettling, doesn’t it? Simular is betting that the future of work involves software that doesn’t just suggest what you should do, but actually takes over your mouse and keyboard to get it done.
Simular is a startup focused on building AI agents that run directly on your Mac and Windows machines, and it recently raised a $21.5 million Series A round led by Felicis, with support from Nvidia’s venture arm, NVentures, existing investor South Park Commons, and several others. Instead of trying to automate everything inside a web browser like many competitors, Simular is aiming at the entire desktop environment—your operating system, apps, and workflows—on both Mac OS and Windows. In practical terms, that means its software can move the mouse on your screen, click buttons, copy and paste data, and repeat the sorts of tedious digital tasks a human user would normally perform manually.
What Simular’s agent actually does
Think of Simular’s AI like a very patient digital assistant that can physically operate your computer the way you do. It can open applications, navigate interfaces, and perform repetitive actions such as copying data from one place and pasting it into a spreadsheet or internal tool. Instead of building a specialized integration for every app, the agent interacts with the graphical interface itself, which makes it more flexible but also raises big questions about reliability and control. But here’s where it gets controversial: if software is clicking and typing on your behalf, how much trust are you willing to give it over your screen?
On the product side, Simular has already released version 1.0 of its agent for Mac OS, making it available to early users who want to automate workflows on Apple machines. At the same time, the company is collaborating with Microsoft to bring a similar agent to Windows. Simular is one of only five companies selected for Microsoft’s Windows 365 for Agents program, alongside Manus AI, Fellou, Genspark, and TinyFish, which signals that Microsoft sees desktop-level agents as a serious direction for enterprise automation. While the team has not committed to a specific launch date for the Windows version, they suggest that it could eventually attract as much or even more user interest than the Mac release.
Who’s behind Simular
A big part of the buzz around Simular comes from the founders’ backgrounds. CEO Ang Li is a continuous learning researcher who previously worked at Google DeepMind, where he met co-founder Jiachen Yang, a specialist in reinforcement learning. Rather than doing purely theoretical research, their work focused on applying advanced AI methods to real-world products inside Google, including projects tied to Waymo, Google’s autonomous driving division. That mix of academic rigor and product-focused experience is crucial when turning futuristic AI ideas—like autonomous agents that run your computer—into something robust enough for business use.
This matters because the dream of “agentic AI”—systems that can break down goals into many steps and handle them with little oversight—runs into a very practical problem: today’s large language models sometimes hallucinate. In this context, hallucination means the AI confidently generates incorrect or made-up information, and when an agent has to execute thousands or even millions of small actions, a single wrong step can corrupt an entire workflow. The more steps an agent takes, the higher the total chance that one of them will be based on a hallucination.
The hallucination problem in agents
Agentic systems often have to chain together long sequences of actions, from reading files to clicking buttons and entering data. If the AI misinterprets a screen, fabricates a result, or chooses the wrong action at any of those points, it can derail a complex process—imagine an agent mis-filing documents, overwriting records, or sending the wrong emails. Because each step adds to the probability of an error, agents that rely purely on non-deterministic language models become increasingly risky as they tackle more ambitious tasks. And this is the part most people miss: “smart” doesn’t always mean “safe,” especially when software has full control over your device.
One classic way to reduce hallucinations is to make the underlying language model behave more deterministically. Instead of allowing the model to generate a wide variety of possible answers, its outputs are constrained so that the same input always leads to the same response. That can make behavior more predictable, but it also risks flattening the model’s creativity and flexibility, which are exactly what make agents powerful for open-ended problem solving. So there’s a tension: lock things down for safety, or keep them flexible for innovation.
Simular’s hybrid approach
Simular claims it has found a middle path between free-form exploration and rigid determinism. The company’s agent initially works in a flexible, exploratory mode, where it tries different approaches to accomplish a task while a human user stays “in the loop” to monitor and correct its behavior. The user can step in as the agent iterates, steering it back on course until it finally completes the workflow successfully from end to end. Once that reliable sequence of actions is established, the user can save or “lock in” that workflow, turning it into a deterministic routine that the agent can repeat without improvisation.
Ang Li describes this as letting the agent explore until it finds a “successful trajectory”—a path through the task that consistently works. After that, the system converts that trajectory into deterministic code, effectively freezing the winning strategy into a repeatable program. In simple terms: first the agent learns by trial and error with your guidance, then it graduates into a stable automation script you can trust to run the same way every time. This design tries to preserve the creative, problem-solving advantage of LLMs during the learning phase while delivering the reliability of traditional software once a workflow is finalized.
Under the hood: “neuro-symbolic” agents
Simular emphasizes that it is not just wrapping a language model with a simple interface and calling it an agent. Instead, the team is building what they describe as “neuro-symbolic computer use agents,” which combine neural network–based AI (like LLMs) with more traditional, rule-like or symbolic structures. According to Li, their system is deliberately not purely LLM-based: the language model’s role is to generate code that then becomes deterministic, executable logic. Once a workflow has been codified this way and proven to work, running it again should produce the same result consistently, because the execution no longer depends on the LLM’s creativity.
A notable benefit of this approach is that the resulting deterministic code lives with the user, not hidden inside the model. Users can open, read, and audit the code that represents each automation, building trust by understanding exactly what actions will be taken and in what order. For more technical teams, this also opens the door to version control, code review, and security checks, much like any other script or internal tool. But here’s where it could be contentious: will average non-technical users actually inspect this code, or will they simply click “run” and assume it’s safe?
Real-world use cases and traction
It is still early days for Simular’s technology, and it remains to be seen whether this method will be the breakthrough that brings AI agents onto every worker’s desktop. Nonetheless, the company already reports several early beta customers using the system in practical scenarios. For example, one car dealership is using the agent to automate vehicle identification number (VIN) lookups, a repetitive but important task in sales and inventory workflows. Homeowners’ associations (HOAs) are using it to pull key contract details out of PDF documents, helping them avoid hours of manual reading and data entry.
Beyond commercial pilots, Simular also offers an open-source project, currently available only for Mac OS users, which has been used to build automations in content creation, sales, and marketing. These range from assembling outreach campaigns to generating and organizing marketing materials. These examples highlight a bigger theme: many office jobs involve repetitive digital tasks across multiple apps, and if Simular’s agent can reliably handle those, it could free teams to focus more on judgment, strategy, and relationship-building.
Funding and future stakes
On the financial side, Simular previously secured a $5 million seed round before closing its recent $21.5 million Series A, bringing its total funding to roughly $27 million. In addition to lead investor Felicis and participants NVentures and South Park Commons, other backers include Basis Set Ventures, Flying Fish Partners, Samsung NEXT, Xoogler Ventures, and angel investor and podcaster Lenny Rachitsky. That investor mix suggests confidence that desktop agents could become a foundational layer of enterprise productivity, not just a niche experiment.
The company’s momentum also lines up with broader industry interest in agentic AI, showcased at major tech events like those held in San Francisco, where the conversation increasingly shifts from “What can AI say?” to “What can AI actually do on my behalf?” Still, the big open question is whether workers and companies are comfortable letting a third-party agent operate their machines in such a direct way. Is the promise of saving hours of manual clicking and typing enough to offset concerns about security, control, and mistakes that might be hard to trace?
Your turn: would you let an AI drive your desktop?
Simular’s vision raises some provocative questions. If an AI agent can reliably automate your repetitive tasks by directly controlling your computer, is that empowering—or is it a step too far in handing over control to software? Would you be comfortable with an AI moving your mouse, clicking buttons, and entering data on your behalf if you could review and audit the underlying code, or does the very idea feel risky regardless of safeguards?
And here’s the part that might divide opinions: is turning exploratory AI behavior into deterministic code really the best way to solve hallucinations, or does it simply mask deeper issues with today’s models? Share your thoughts—do you see this “neuro-symbolic” approach as the future of work automation, or are you skeptical about letting agents run wild (even briefly) on your personal or corporate machines?
