Simular’s AI Agent Wants to Run Your Mac and Windows PC for You

Simular, a startup developing AI agents for macOS and Windows, has raised $21.5 million Series A led by Felicis, with participation from NVentures (Nvidia’s investment arm), South Park Commons, and others.

What makes Simular notable is that, unlike many other agentic AI companies focused on browser-based automation, Simular aims to control the entire computer system. Agentic AI refers to systems capable of autonomously performing complex tasks with minimal human involvement.

“We can literally move the mouse on the screen and do the click,” co-founder and CEO Ang Li told TechCrunch, explaining that the agent can repeat almost any digital task a human performs — such as copying and pasting data into a spreadsheet.

On Monday, the company announced the launch of Simular 1.0 for macOS. Simular is also working with Microsoft to develop a Windows version. The startup was accepted into Microsoft’s newly announced Windows 365 for Agents program, alongside Manus AI, Fellou, Genspark, and TinyFish. Li didn’t provide a release date for the Windows version but suggested it could become as popular — if not more — than the Mac release.

Simular’s founders bring strong credentials. Ang Li previously worked as a continuous learning researcher at Google DeepMind, where he met co-founder Jiachen Yang, a reinforcement learning specialist. Their work at DeepMind wasn’t strictly academic, Li said — it directly contributed to improving products such as Waymo.

The Technical Hurdle: LLM Hallucinations

Before agentic AI can become mainstream, several technical challenges need to be solved — the biggest being LLM hallucinations. Even a small error rate becomes disastrous when an AI agent must complete thousands or millions of steps. A single hallucinated action can derail the entire workflow.

Some companies attempt to solve this by making LLMs deterministic — ensuring outputs are the same every time. But this sacrifices the creativity and flexibility that make agents useful in the first place.

Simular’s Hybrid Approach: Exploration First, Determinism Later

Simular combines both approaches. Its agent explores tasks under human oversight, iterating freely until it succeeds. Then, the user can “lock in” that successful process.

“Once you’ve found a successful trajectory, that becomes deterministic code,” Li explained. The result is a workflow that is reliable, repeatable, and stable.

Simular can achieve this because its system is not just an LLM wrapper.

“We use a new technology that’s not used by any other agent company. We call it ‘neuro-symbolic computer use agents.’ It’s not fully LLM-based,” Li said. “Our approach is to let the LLM write deterministic code. If a workflow works once, it will work every time.”

An added benefit: the deterministic code is visible to the user.
“Once they have the code, they can trust it because they can inspect it, audit it, and understand exactly what’s happening,” Li said.

Early Use Cases and Open Source Adoption

Simular’s early beta customers come from varied industries:

• A car dealership is using the agent to automate VIN searches
• Homeowners Associations (HOAs) extracting contract information from PDFs
• Creators using Simular’s open source project for content generation
• Sales and marketing teams are automating repetitive workflows

The open source version is currently available only for macOS.

Simular previously raised a $5 million seed round, bringing total funding to roughly $27 million. Additional backers include Basis Set Ventures, Flying Fish Partners, Samsung NEXT, Xoogler Ventures, and angel investor Lenny Rachitsky.