Simular, a company developing artificial intelligence agents for both Mac OS and Windows operating systems, has secured $21.5 million in Series A funding. The investment round was spearheaded by Felicis, and also included participation from NVentures, Nvidia’s venture capital division, South Park Commons – a previous seed investor – and several other firms.  

Simular distinguishes itself within the growing field of agentic AI by focusing on controlling the personal computer directly, rather than simply managing web browser activity. (Agentic AI describes systems capable of independently executing intricate tasks with minimal human oversight.) According to co-founder and CEO Ang Li, in a discussion with TechCrunch, the agents possess the ability to physically manipulate on-screen elements, such as moving the mouse and performing clicks, enabling them to replicate a wide range of digital human actions, citing the example of transferring data into a spreadsheet. 

The company announced the release of its version 1.0 for Mac OS on Monday. Simultaneously, Simular is collaborating with Microsoft to create an agent specifically designed for Windows. The startup is one of five agentic companies selected for the Windows 365 for Agents program, which Microsoft unveiled in mid-November. (The other participating companies are Manus AI, Fellou, Genspark, and TinyFish.) Regarding the release timeframe for the Windows version, Li remained noncommittal, but indicated expectations for its popularity to equal or surpass that of the Mac OS version. 

The strength of Simular’s founding team is another compelling reason to observe its progress. Li is a specialist in continuous learning, with prior experience at Google’s DeepMind, where he connected with his co-founder, Jiachen Yang, a reinforcement learning expert. Li clarified that their team’s research wasn’t purely theoretical; it was geared towards enhancing Google’s products, including Waymo. 

This background in AI product development is advantageous, as numerous technical challenges must be addressed before the envisioned agentic future becomes a reality. A significant hurdle is the tendency of large language models (LLMs) to generate inaccurate or misleading information – often referred to as “hallucinations.”

Successfully completing agentic tasks can necessitate the execution of thousands or even millions of individual steps. A single hallucination at any point in the process can invalidate the entire agent’s work, and the likelihood of such errors increases with the number of steps involved.  

One potential solution involves transforming the “non-deterministic” nature of LLMs into a “deterministic” one, meaning restricting the LLM’s creative freedom to ensure consistent responses and actions. However, this approach could limit the agent’s capacity for innovative problem-solving.

Simular is adopting a combined strategy. Its agent will explore various approaches to a task, with the human user providing guidance and corrections until a successful outcome is achieved. The user can then solidify that task’s workflow, making it deterministic and repeatable. 

“Our method involves allowing agents to continue exploring successful paths. Once a successful path is identified, it is converted into deterministic code,” Li explained. 

The startup’s ability to implement this approach stems from its technology, which, as Li acknowledges, is still in its early stages, and goes beyond a simple LLM interface for sending and receiving data.

“We have developed a novel technology that is not currently utilized by any other agent company. We refer to it as ‘neuro symbolic computer use agents.’ It is not solely based on LLMs,” he stated. “Our approach to mitigating hallucinations is to have the LLM generate code that becomes deterministic. Therefore, if a workflow proves effective, subsequent executions of the same workflow will also be successful.”  

A further advantage is that this deterministic code, which performs the repeatable task, resides with the end user, not the LLM. “Users can have confidence in the code because they can examine it, audit it, and understand its operation,” Li added.

Whether this methodology will be the key to widespread agent adoption remains to be seen. Li reports that initial beta customers include an automotive dealership automating vehicle identification number (VIN) searches, and homeowner associations extracting information from PDF documents. Furthermore, the company’s open-source project (currently available only for Mac OS) has facilitated automations in areas such as content creation, sales, and marketing.  

Simular had previously raised $5 million in seed funding, bringing its total funding to approximately $27 million. Additional investors in the company include Basis Set Ventures, Flying Fish Partners, Samsung NEXT, Xoogler Ventures, and angel investor and podcast host Lenny Rachitsky, according to the company.