The small humiliation of applied artificial intelligence is that reality gets a vote. A model can look brilliant on Tuesday and stumble on Wednesday because a page changed, a new query arrived or a familiar button moved six pixels to the left. Ang Li learned this before AI agents became a category crowded with launch videos. At Google DeepMind, where he worked from 2017 to 2021, production engineers had a blunt refrain: machine learning never works in production.
Li saw the joke and the warning. Research papers promised astonishing capability, yet live systems were full of moving parts. He recalls an attempt to apply an AlphaGo-style system to Google's advertising operation. The result was not a triumphant revenue chart. Revenue dropped. The real system, he concluded, was simply too complex for the comforting assumption underneath much machine learning: that the data stays put.
YouTube receives new videos. Search gets new queries. A road acquires new hazards. The world redraws the test while the model is taking it. Li's career can be read as a long argument with that instability - first through perception research, then maps and ads, then autonomous driving, and now the ordinary desktop.
“Our definition for agents is a system that can interact with the environment and keep improving itself.”Ang Li on what separates an agent from an assistant
A researcher follows the moving target
Li started coding early enough for it to become a competitive sport. In 2013, he was part of the team that won the Mid-Atlantic regional ACM/ICPC contest. He later reached the world finals and placed 39th. Competitive programming rewards speed, but it also rewards something more durable: turning a messy prompt into an exact sequence that the machine can execute.
At the University of Maryland, College Park, he pursued the messier side of the equation. His 2017 computer science doctorate focused on representation learning for visual and linguistic information, with a thesis on spatial perception. The question was how a machine could connect words to the visible world. His publication record widened from computer vision and model efficiency to reinforcement learning, uncertainty, security and continual learning.
The list of places on his resume reads like a tour through the practical surfaces of AI: Comcast Labs DC, Google Street View, Apple, Carnegie Mellon Robotics, Facebook AI Research, Google DeepMind and Baidu Apollo. The connecting tissue is not a single product. It is the effort to make learned systems behave somewhere other than a sealed laboratory.
At DeepMind, Li led applied programs involving evolutionary neural networks and lifelong learning. The work reached systems people use without thinking about the machine intelligence underneath them. His publication credits include graph neural networks for predicting arrival times in Google Maps and work on video understanding. From 2021 to 2023, as a principal scientist at Baidu Apollo, he led continual-learning platform development for self-driving cars.
A car offers a severe version of the changing-world problem. A system cannot politely ask the road to resemble its training set. It must perceive, choose and adapt while the environment keeps arriving. Continual learning is the attempt to make new experience useful without causing a model to forget what it learned before. Li calls that forgetting problem one of the hardest in AI.
The interface was already here
Li co-founded Simular with Jiachen Yang in 2023. They were not strangers meeting over a fashionable market map. The two had previously published work together on learning to incentivize other learning agents. Yang became Simular's chief technology officer; Li became chief executive. Their company started with an idea that then required explanation: they were building agents.
People suggested a softer word. Call them assistants, they said. Agent sounded unfamiliar. The naming debate now feels like a message preserved from another geological layer of AI. Within two years, nearly every model company and a parade of startups had adopted the agent label.
Li's definition remains narrower than the marketing category. An agent must act in an environment and improve. A chatbot can explain how to complete an expense report. A computer-use agent can open the relevant software, move through the fields and finish the workflow. It meets software where humans already do: at the graphical interface, through windows, menus, clicks and keystrokes.
The architecture behind Simular's research tries to prevent every task from becoming a fresh improvisation. A large language model is useful for exploration: inspect a screen, consider paths, try an action. Once the system finds a solution, it can convert that trajectory into symbolic code and execute it predictably. When the code breaks because the environment has changed, the model returns.
This is a pragmatic split. Models handle ambiguity; programs handle repetition. The agent earns a routine, then stops paying the cost of rediscovering it each time. Reliability becomes a memory of success rather than a hope that a generative model will have the same good idea twice.
“Agents must be able to reliably use a computer in real environments, not just in simulations.”Ang Li at the launch of Simular 1.0
An open framework, in public
Simular made its research legible through Agent S, an open framework released with code in 2024 and published at ICLR in 2025. The system combined knowledge from external search with memories of prior experience, planning tasks as hierarchies instead of one long guess. Agent S2 followed with a team-of-models approach: generalists for broad reasoning, specialists for grounding actions precisely on a screen.
Later versions pushed performance further, but the progression matters more than a frozen benchmark score. Each generation isolates a reason agents fail - poor grounding, brittle long-range plans, uneven model skills, unreliable repetition - and treats it as an engineering problem. The work is public enough for developers to inspect, run and challenge.
The company also needed a product language less severe than benchmark tables. Simular introduced Bigfoot, a talking character that appeared in its marketing and then inside Simular 1.0 to offer tips and flag problems. Li has discussed the decision with a grin. It is a small admission that human-computer collaboration is partly social design. A system that asks for correction needs a way to feel correctable.
The business of repeatability
In December 2025, Simular announced a $21.5 million Series A led by Felicis, with NVentures, Basis Set Ventures, Flying Fish Partners, South Park Commons and Lenny Rachitsky participating. Basis Set had led a previous $5 million seed round. The new capital arrived alongside Simular 1.0, a native desktop agent designed to work across applications, handle longer workflows and accept natural-language redirection while it runs.
Li positions Simular as an infrastructure company, not merely another agent product. The ambition is a neuro-symbolic continual reinforcement-learning framework that can underwrite many agents and many devices. In his telling, today's dependence on carrying and operating computers is temporary. The computer itself could become an active participant that books, organizes, fills and follows through.
That future introduces a delicate standard. Capability attracts attention; reliability earns access. The closer an agent moves to real documents, company systems and expensive decisions, the less charming a confident mistake becomes. Li argues that reliability matters more than a growing list of isolated abilities. The machine needs to know how to cooperate, accept a correction and repeat a success.
No permanent walls
Li's founder philosophy carries the same expectation of change. Asked about defensibility in AI, he has offered a wonderfully inconvenient line: “There's never a moat on anything.” The underlying technology shifts too quickly for a founder to treat one technical lead as permanent property. A model provider can ship a feature; a platform can redraw the boundary; a rival can catch up.
His answer is motion rather than fatalism. Keep the team small enough to decide. Learn from customers quickly. Choose which discoveries should become infrastructure and which assumptions should be discarded. The company's architecture and the founder's operating theory rhyme: explore, preserve, adapt.
In July 2026, Li took that thesis to Seoul during ICML week, hosting a researcher roundtable and discussing Korea's possible role in the agent era. His argument extended beyond software. If agents require a body, Korea's strengths in chips, memory, phones, displays and appliances could matter. The autonomous computer, in this view, is not one app. It is a rearrangement of who operates the world's devices.
The horizon remains grand. Li's personal mission is to build accessible general intelligence that liberates people from repetitive computer labor. Yet the path he describes is insistently concrete: a mouse movement, a keyboard action, a workflow completed, a correction remembered. General intelligence arrives, if it arrives, through contact with particular things.
That may be the most useful part of Li's story. He is not waiting for the world to become clean enough for intelligence. He is building for the world that changes the page, uploads the next video, redraws the road and asks the machine to try again.