Simular’s AI Agent Can Run Your Mac or Windows PC
▼ Summary
– Simular, an AI agent startup, has raised $21.5 million in Series A funding led by Felicis, with participation from NVentures and others.
– Its unique approach involves controlling the PC operating system directly to automate human-like digital tasks, rather than just the browser.
– The company is developing agents for both Mac OS (with a 1.0 release) and Windows, being part of Microsoft’s Windows 365 for Agents program.
– Founded by ex-DeepMind scientists, Simular addresses the hallucination problem in LLMs by converting successful agent workflows into deterministic, auditable code.
– This “neuro symbolic” method allows users to lock in and trust repeatable automations, with early applications in areas like data extraction for car dealerships and HOAs.
A new startup called Simular has secured significant funding to develop a unique type of AI assistant that directly controls your computer’s operating system. The company recently closed a $21.5 million Series A funding round led by Felicis, with participation from Nvidia’s venture arm, NVentures, and existing investor South Park Commons. Unlike many AI agents that operate within a web browser, Simular’s technology is designed to take direct control of a Mac or Windows PC, performing tasks by moving the cursor, clicking, and typing just as a human would.
This approach allows the AI to replicate a wide range of digital activities. Co-founder and CEO Ang Li explained that the agent can handle complex, multi-step workflows, such as copying data from one application and pasting it into a spreadsheet. The startup has just launched its 1.0 version for Mac OS and is also collaborating with Microsoft as part of a select group of companies in the Windows 365 for Agents program. While a release date for the Windows version isn’t specified, Li anticipates it could become even more popular than the Mac iteration.
The founders bring considerable expertise to the challenge. Li is a continuous learning scientist, and his co-founder, Jiachen Yang, is a reinforcement learning specialist. Both previously worked at Google’s DeepMind, where their research was applied to real-world products like Waymo. This practical experience is crucial for tackling one of the biggest hurdles in agentic AI: the tendency for large language models (LLMs) to hallucinate, or generate incorrect information. In a long sequence of actions, a single error can render an entire automated process useless.
Simular’s proposed solution to this problem is a hybrid method. The AI agent initially explores different ways to complete a task, with a human user providing guidance and corrections. Once a successful method is found, the user can “lock in” that specific workflow. The system then converts the successful steps into deterministic, repeatable code. This process aims to balance the AI’s creative problem-solving with the reliability of a scripted program.
The company’s core technology, which it calls “neuro symbolic computer use agents,” is not solely reliant on an LLM. Instead, the LLM is used to generate code that executes tasks deterministically. This means that after a workflow is successfully established, it will perform identically each time it is run. A significant advantage of this system is that the resulting code is owned and controlled by the end-user. They can inspect, audit, and modify it, building trust in the automation.
Early use cases for Simular’s beta software include a car dealership automating vehicle identification number searches and homeowners’ associations extracting data from PDF contracts. The company also maintains an open-source project that has inspired automations for content creation and sales tasks. With a total of roughly $27 million in funding from investors like Basis Set Ventures and Samsung NEXT, Simular is positioning its unique code-generating approach as a potential key to bringing reliable, user-controlled AI agents into mainstream business operations.
(Source: TechCrunch)
