Yu Su: From Logical Agents to Language Agents and Specialized Learning

Zhang Xiaojun invites Ohio State University computer science professor and NeoCognition founder Yu Su to discuss the technical evolution of agents. Su studies language agents and computer use, and traces the field from early logical systems to large language models and personal agents.

Su studied computer science at Tsinghua, earned a PhD in the US, and established an NLP group at Ohio State. Its work includes Mind2Web, LLM Planner, computer-use agents, and multimodal evaluation. He later founded NeoCognition to continue agent research.

He offers a broad definition of an agent: a bounded entity that acts in an external environment toward goals rather than wandering randomly. Under that abstraction, people and other animals can also be considered agents.

Building artificial agents that imitate people was a goal from AI’s early days, Su says. Because complete agents were so difficult, the field split into areas such as vision, natural language processing, and logical reasoning; recently these directions have begun converging again.

He calls an early representative the logical agent: expert systems encoded domain knowledge as logical rules and used inference engines to answer new questions. Their expressive limits and the cost of acquiring knowledge eventually prevented them from fulfilling many expectations.

To analyze agents, Su proposes two broad dimensions: memory and autonomy. Memory includes representing, acquiring, updating, and forgetting knowledge, including semantic, episodic, and procedural knowledge. Autonomy includes perception, reasoning, decision-making, and actions affecting the environment.

The two depend on each other: agents use existing knowledge to perceive and act and can update it from action. Early logical systems’ memory was bounded by their rule language, restricting their reasoning and actions; transcribing expertise into formal rules was particularly laborious.

He recommends Stuart Russell and Peter Norvig’s Artificial Intelligence: A Modern Approach as an introduction to early AI and says its authors regarded agents as a central theme. Agents are not a concept invented only recently.

With neural networks and deep reinforcement learning, researchers trained agents to play Go, arcade games, and other complex games. Su says these systems made substantial progress in particular settings, but inputs, outputs, memory, and actions were generally constrained by the task.

He explains a limitation of early neural agents through computation: some produced an action from an image in one forward pass, with reasoning largely implicit in the network, while humans may spend different amounts of time and computation on problems of different difficulty. He uses this contrast to examine generality.

Compared with logical agents, neural methods can apply similar architectures and training approaches across games, but that does not mean an agent placed in any new environment can learn as quickly as a person. Early deep reinforcement learning often required many repeated samples, he says.

Why did games become important experimental environments? He suggests that beyond researchers’ interests, games offer repeatable, abundant virtual interaction and relatively clear inputs and actions. He did not work directly on that type of research, so presents this as a limited explanation.

Meanwhile, NLP developed semantic parsing: converting human language into machine-processable forms for querying databases or knowledge graphs or operating websites. Su considers this complementary to neural agents in constrained environments, expanding language understanding and expressible actions.

He studied semantic parsing during his PhD and later moved into language models and agents alongside others with similar backgrounds. Language allows people to express tasks flexibly and can organize reasoning, represent knowledge, and connect tools.

Su calls the new generation based on large language models “language agents.” Language can scaffold perception, reasoning, and action; systems may use different lengths of reasoning for different tasks rather than applying the same computation each time.

Looking back, he cites his team’s LLM Planner and Mind2Web work around 2022, exploring language models for robot planning and website use. An important contribution may look simple in form, he says, but recognizing and implementing a viable direction at the right time is difficult.

He also mentions early tool-use work and AutoGPT, which demonstrated possibilities for models to call external tools and organize actions, though early systems could reliably complete few tasks. GitHub popularity signals community interest, not necessarily mature capability.

As multimodal models improved, web agents began relying more on visual information. Su describes his team’s multimodal evaluation and visual web-agent work, alongside other teams’ efforts to build reproducible web-task environments.

He compares live sites and simulated environments: live websites are closer to the open world but change and may raise permissions and legal issues; controlled simulated sites are more reproducible. Both have uses, and results need to be understood in their environment.

Research later expanded to desktop, mobile, and coding. Su describes UGround’s approach: an agent observes the screen and acts through clicks and typing like a person rather than relying solely on a website’s underlying text structure.

He observed rapid diffusion of coding agents in 2025 as foundation models improved at programming. The spread from Silicon Valley to other regions still takes time, he says; his timeline and assessment reflect personal observations from the interview.

Web, desktop, mobile, graphical interfaces, command lines, and APIs are different means of action. Such distinctions have historical uses, Su says, but boundaries are changing, and the long-term goal may be a general agent able to perform many tasks in the digital world.

He particularly values coding because digital systems are built and operated through code, which can connect interfaces and tools. He still uses “language agent”: programming languages are languages too, and language describes the world, communicates, and enables action.

His move from semantic parsing to language agents also grew from concern about the relationship between people and software: complex software often makes people spend substantial time learning how machines work. He wants machines to better understand human expression so people need not always think like machines.

He compares the OpenClaw moment with ChatGPT’s adoption: underlying capabilities had developed for some time, but a new interaction let more people experience what they could do. It brought personal agents to wider attention, he says, without erasing earlier research.

Giving agents more permissions can broaden tasks but also create potential harm. Messaging interfaces and always-on environments change the experience, but permissions, safety, and control still require care; open source alone does not eliminate risk.

Su thinks OpenClaw’s popularity may have accelerated companies’ attention to agents and productivity tools. His social-impact assessment is a prediction from the interview; specific product strategies and employment changes cannot be attributed to one project alone.

Comparing diffusion in China and the US, he observed that US discussion was then more concentrated among developers and technology communities, while in China it reached a broader public. Faster application adoption may be an advantage, he says, but that does not mean all users have already found value.

Improved model capabilities may reduce friction and make previously uneconomic tasks worthwhile, he says. Discovering genuine needs, designing products, and executing still matter, and hype-driven spending and waste can occur.

He worries that society could face pressure if jobs are displaced faster than new opportunities appear. Researchers should make agents more reliable and easier to use, he says, so more people can turn their ideas into practice rather than only a few technical teams benefiting.

Asked whether only model companies can capture product opportunities, Su says broad platforms may suit large model providers, but industries, companies, and workflows contain specialized knowledge. Helping agents adapt to these different settings leaves room for other teams.

He describes NeoCognition’s goal as agent research with particular attention to adapting general capabilities to specific work. Each occupation, company, and software environment is a “small world” with rules and context to learn; specialization is more than adding a generic prompt.

The company hopes to study cross-domain methods, with enterprise work likely a focus because such tasks often require deep expertise. Su does not think all software will be rewritten or disappear; existing vendors are also trying to move from providing tools to delivering more outcomes.

He says NeoCognition raised approximately $40 million in seed funding and describes the process as relatively smooth. In his observation, leading projects and ordinary teams may face different financing conditions in the US; market percentages mentioned in the interview are personal estimates.

Investors consider team experience, differentiation, feasibility, and potential commercial value, and ask whether large model companies can easily copy the idea. Su believes genuinely difficult, high-potential problems may accommodate multiple technical approaches, though that does not guarantee any company’s success.

He compares benchmark design with choosing a startup problem: a good benchmark should measure capabilities related to real-world value rather than merely optimize an easy-to-count number. He wants to research ways for general agents to learn specialized tasks faster, while acknowledging that the problem is unsolved.

Su uses a broader definition of world models than visual generation alone. Besides predicting images, a model can represent organizational relationships, workflows, software rules, and interactions between people: aspects of an environment an agent may need to understand for work.

He uses a new employee as an example: an intern gradually learns the company’s organization, approvals, tools, and colleagues, developing an understanding of the workplace. He describes the path from novice to expert as continually building and revising a model of that “small world.”

Current general agents can sometimes produce plausible results but may not be sufficiently reliable, fast, or economical, he says. Specialized learning aims to improve reliability and efficiency through experience in a particular environment; this is a research goal, not a solved capability.

“Continual learning” has multiple meanings, he notes. A classic machine-learning problem is learning new tasks without forgetting old ones; the term is also used for personalization, model updates, and other processes, so the object and goal of learning need clarification.

He personally wants to link continual learning with a broad world model: a system accumulates understanding of its environment to act better. This is his research framework, he emphasizes, not a definition shared by all researchers.

He discusses human learning through research on the neocortex and cites Jeff Hawkins’s A Thousand Brains as one theoretical perspective. There are competing hypotheses about how cortical structures form world models; the analogy should not be treated as a fully established neuroscience conclusion.

On language and world models, he considers vision and other perception important but says language and symbols also help people communicate, accumulate, and transmit knowledge. He cites views on symbols and human evolution as an ongoing research question.

He mentions research distinguishing language processing from individual thought and offers his interpretation: language may scaffold learning, while internalized knowledge may not require language at every step of thought. Symbolic expression remains important to social cooperation and knowledge transmission, he believes.

Even if future agents are based on something beyond large language models, he expects language to remain important for understanding, reasoning, and interaction. Whether the name retains “language” may change as the capability becomes commonplace.

Graphical interfaces and command lines are both means to an end, Su says. GUIs remain useful for people to inspect, verify, and audit results; agents using existing interfaces can benefit from business processes, constraints, and knowledge already encoded in them.

Many long-tail applications have GUIs but may not quickly rewrite APIs or command lines for agents. Visual representations may also suit some complex relationships, so he does not expect command lines to replace GUIs entirely.

He cites the semantic web to illustrate the difficulty of adopting new standards: even if a unified machine-readable interface is useful in some circumstances, participants need sufficient incentives and resources to change existing systems. Keeping a local solution can be more sensible than rewriting everything.

He connects memory, continual learning, world models, and specialization in his account of the bottleneck: continual learning is the process, an environmental model is the content, and an expert agent is the goal. The intended improvements include reliability, speed, and cost.

He expects continual learning to be an important agent research direction in 2026, with multiple possible approaches. Enterprise deployments still require engineers to help build workflows, reflecting unresolved reliability, cost, and accessibility problems.

Comparing companies’ agent efforts, he sees some convergence in product strategies alongside different experiments in visual interaction, coding, and enterprise work. His descriptions of internal strategies and market positions are observations and speculation from the interview.

Asked about long-task reliability and safety, Su argues that inexperience and ineffective learning can lead to mistakes, and familiarity with an environment may reduce some risks. He distinguishes safety issues from security against worst-case scenarios, which still requires dedicated methods.

He explains his move from academia to entrepreneurship: university let him explore many ideas simultaneously, but larger experiments and deployments increased the need for compute, capital, and an execution team. He presents this as personal experience, not a view representing all academics.

Starting a company brought a new pace. Su enjoys connecting knowledge into conceptual frameworks and finds new problems in entrepreneurship. He also acknowledges that startups need market validation, customers, investors, and product-market fit, creating incentives different from academic research.

He sees some overlap between managing a university lab and running a startup: leaders secure resources, recruit and develop people, and drive results. Academia generally offers more research freedom, though funding pressure and institutional context matter.

Recalling childhood, he says he loved games and read widely. Reading let him construct different worlds in his mind and find connections between them, a habit he believes shaped his learning and cross-disciplinary frameworks.

He grew up in a small Hunan county and recalls concentrating his preparation before important exams. Tsinghua’s independent admissions offer gave him reassurance; he ultimately entered on his college entrance exam score, and a scholarship eased the burden on his family.

He describes a combination in his personality: he does not want to treat every outcome too heavily, yet puts effort into things he truly wants to do. He says this has influenced his research, teaching, and entrepreneurship.

Agent research and deployment are increasingly difficult to separate, he believes: learning continually from real use first requires access to real environments. He says his startup timing depended on whether the technical foundations and his own understanding were ready.

He predicts progress in continual learning over the next few years that may support wider agent adoption. This is his research judgment; the pace, capabilities, and social effects remain uncertain.

His main immediate social concerns are job displacement and distribution of gains: if new jobs are insufficient, gains concentrate, and effective redistribution is absent, technical progress could cause major disruption. His personal view on longer-term risks does not establish that such risks are ruled out.

Researchers can work to lower barriers to advanced agents so more people with expertise and ideas can build tools and create value, he says. Achieving broad access and fair distribution also requires work beyond technology.

Finally, Su recommends A Brief History of Intelligence and A Thousand Brains. The former connects AI, evolution, and neuroscience and inspired some of his research; he calls the latter a bold theory of the brain while acknowledging the evidence is early.

On influential AI research, he ranges from early computational neuron models, backpropagation, AlexNet, and word embeddings to attention, Transformer, BERT, GPT, chain-of-thought, and tool use. He emphasizes continuity: new methods often build on prior work.

Asked for his most important current research bet, he again chooses continual learning and world modeling. He also clarifies that his upbringing is a personal story, not encouragement for young people to emulate intensive gaming or skipping school.