Ang Li

Ang Li is the co-founder and CEO of Simular, where he builds agents that operate computers through the applications people already use. His work connects continual learning with a concrete problem in automation: how to preserve a successful workflow so an agent can repeat it, while still adapting when the interface or task changes.

From perception to autonomous computers

Ang earned his Ph.D. in computer science at the University of Maryland, College Park, studying spatial perception from visual and linguistic information. His research background spans computer vision and learning systems. During internships at Apple and Facebook AI Research, deploying deep learning exposed him to the distance between research results and working products. At DeepMind, he built a prototype that developed into a collaboration with Waymo. He subsequently led self-driving work at Baidu before leaving in 2023 to start Simular.

In his account of that transition, Ang connects autonomous driving with his interest in systems that keep learning. Computers offered an environment where experiments and improvements could proceed without a vehicle’s manufacturing cycle. That interest preceded the generative AI boom.

Two earlier collaborative research projects show different aspects of his work on learning systems:

• Adapting training as it runs: As first author of a generalized population-based training framework, Ang and his collaborators developed a system that distributes training trials asynchronously across a cluster. Trials copy successful model weights and change hyperparameters during training. Its black-box design accommodates different model architectures and training procedures.
• Learning without forgetting: Ang co-authored Orthogonal Gradient Descent, which addresses catastrophic forgetting: learning a new task can damage a network’s performance on earlier ones. The method constrains new gradient updates to preserve previously learned outputs while allowing further learning.

Ang founded Simular with fellow former DeepMind researcher Jiachen Yang in 2023. The company’s first browser-agent release, in April 2025, used a native macOS application with an embedded browser, allowing users to intervene while the agent worked. Simular subsequently expanded to Sai, an agent for local and cloud computers, alongside developer access through an autonomous-computer API.

Planning, grounding and repeated attempts

Ang co-authored the Agent S series with researchers including Saaket Agashe, Jiachen Yang and Xin Eric Wang. The work addresses several distinct reasons a computer-use agent can fail during a long task:

• Experience-augmented planning: The original Agent S combines external knowledge search with retrieval of earlier experience. It breaks complex work into subtasks and uses retrieved information to guide planning and execution, supplying knowledge beyond what is immediately visible on screen.
• Generalist and specialist models: Agent S2 divides responsibilities across models. Its grounding system locates interface elements, while its planner revises actions as observations change. This separates deciding what to do from identifying precisely where to click.
• Selecting better attempts: In Agent S3’s scaling research, the team explores multiple executions and selects among them using Behavior Judge. The judge compares descriptions of the agents’ behavior; benefiting from additional attempts therefore depends on recognizing which attempt is better.

Ang’s co-authored reliability study asks a further question: why can an unchanged agent complete a task once and fail the next time? It examines execution randomness, ambiguous instructions and behavioral variability. The practical consequences include testing repeated executions and allowing an agent to clarify what a user means, rather than treating one successful run as sufficient evidence of dependable performance.

Reusing successful work without freezing it

Ang argues that recurring office procedures cannot economically require a frontier model to reconsider every keystroke and mouse click. Simular’s neurosymbolic approach uses models to discover workflows, stores repeatable procedures as code, and independently checks the resulting screen state. Reusable code reduces repeated model reasoning; model-driven exploration provides a way to recover when the stored procedure no longer works.

A computer-use workshop demonstration makes that tradeoff visible. An agent finds a speaker on Tuesday’s conference schedule and saves the successful steps as a skill. When asked to repeat the task for Wednesday, a hiccup sends it back into model-driven exploration before it completes the workflow. The example captures both the value and the limitation of retaining experience: a saved procedure can spare repeated discovery, but it still needs checks and a recovery path when circumstances change.