Simular: Computer-using agents for real-world workflows

Despite how much of life has moved to mobile, most knowledge work still happens on the desktop. People spend the day clicking, typing, and working across multiple apps. And as much as we talk about agents helping us work faster, AI has mostly lived inside chatbots or, more recently, browser-based helpers, disconnected from most of the world’s workflows.

Computer-using agents change that. Instead of manually executing every step, you describe the outcome you want, and the agent completes the work across the software you already use, remembering it for next time. Automating the mundane elevates you to higher-leverage work, letting you focus on goals and results instead of mechanics.

But making computer-using agents fast and reliable is hard. It’s easy to build a model that works some of the time, but enterprise workflows require determinism and span inconsistent interfaces, legacy systems, and unpredictable environments. It’s a complexity that tends to break most computer-using agents locked into rigid processes. 

Getting an agent to behave like a capable knowledge worker requires a novel approach to both the model and the product.

We’re excited to back the researchers behind Simular, a team training a computer-using agent and building agents that can operate software in real-world knowledge work environments on your behalf.

The founding team has a unique combination of deep research expertise and an instinct for what actually matters to businesses. Ang Li and Jiachen Yang are ex-Google DeepMind and have established themselves as two of the leading researchers on adaptive AI, multimodal learning, and symbolic regression. Like the best researcher founders, they pair their technical brilliance with commercial intuition.

Three things stood out to us that gave us conviction on Simular: 

1. Near human-level performance. OSWorld is a leading benchmark for evaluating an AI agent’s ability to do real tasks on the computer, using desktop apps in addition to a browser, completing cross-app workflows, file operations, and more. Humans score 72% on this benchmark. Simular’s agent is already performing at 69.9%, far ahead of other labs’ models and increasingly reliable enough to take over workflows currently performed by people.

2. Understanding that real-world businesses run outside the browser and on Windows. When companies build in Silicon Valley, it’s easy to lose perspective on broader market realities. We love our MacBook Pros and web apps, but two-thirds of the computers in the real world run on Windows. A lot of economically valuable workflows happen across multiple legacy Windows apps. Simular is as focused on Windows as on Mac, constantly tuning the model based on customer problems and feedback.

3. Unique approach to reliability and determinism. For computer-using agents to be useful in the work context, it needs to be both deterministic and adaptable. The agent needs specific instructions to be precise, but not too specific so it can’t adapt to a new UI update. Executable code is too rigid, confined to action-level instructions. Pure natural language is too vague. In order to make agents reliable and adaptable, the founders created Simulang, an abstraction layer that captures and translates the intention behind an action without being brittle. It’s goal-oriented rather than step-oriented, making it more robust.

Simular is iterating rapidly and building faster than much larger teams, focused on shipping the most reliable, repeatable, and accurate computer-using agents. We’re excited to lead Simular’s Series A and partner with Ang, Jiachen, and the entire team to bring computer-using agents to real environments and real problems.