NeoCognition
Last reviewed
Jun 3, 2026
Sources
8 citations
Review status
Source-backed
Revision
v1 · 1,631 words
Improve this article
Add missing citations, update stale details, or suggest a clearer explanation.
Suggest edit
CtrlK
Last reviewed
Jun 3, 2026
Sources
8 citations
Review status
Source-backed
Revision
v1 · 1,631 words
Add missing citations, update stale details, or suggest a clearer explanation.
NeoCognition is a Palo Alto, California artificial intelligence research company building self-learning AI agents that the founders describe as "specialized intelligence." The company emerged from stealth on April 21, 2026 with a $40 million seed round co-led by Cambium Capital and Walden Catalyst Ventures, with participation from Vista Equity Partners. [1][2][3] It was founded in 2025 by Ohio State University computer scientist Yu Su together with Xiang Deng and Yu Gu, three researchers who had worked together in Su's AI agent lab. [1][4] NeoCognition's stated goal is to build expert agents that learn on the job through continual learning, specializing into domain experts the way a human professional gains expertise over a career. [2][4]
The company should not be confused with the Neocognitron, an unrelated 1980 neural network developed by Kunihiko Fukushima that introduced hierarchical feature detection and is often cited as a precursor to the modern convolutional neural network. [5] NeoCognition is a 2026 commercial startup with no connection to that historical model.
NeoCognition develops AI agents intended for enterprise AI deployment, where reliability and domain knowledge matter more than general conversational ability. The core idea is that a single general-purpose agent should be able to specialize on its own, rather than requiring a team of engineers to hand-build a separate system for each vertical. [2][4] Su frames the approach around how people actually acquire skill. "For humans, our continued learning process is essentially the process of building a world model for any profession, any environment," he told TechCrunch. [4] NeoCognition applies that same model-building process to software agents: an agent observes the structure, workflows, and constraints of the environment it operates in, then builds an internal model of that work and refines it over time. [1][2]
The company argues this is a meaningful departure from how most agents behave today. Many current agents are effectively static once deployed, or they depend on manual updates and prompt engineering to handle new situations. [2] NeoCognition's pitch is that an agent which keeps learning after deployment becomes faster, cheaper to run, and more accurate at a given task as it accumulates experience. [1][4]
The motivation is partly a reaction to how unreliable agents still are. Su has pointed out that leading agents, including coding and computer-use tools from companies such as Anthropic and Perplexity, complete their intended task only around half the time. [4] He put it bluntly: "AI today is fundamentally unreliable when it comes to executing real work that requires deep expertise," and added that with current systems, "every time you ask them to do a task, you take a leap of faith." [1][4] That error rate is tolerable for low-stakes work but not for the high-value tasks enterprises most want to automate, and closing the gap is the problem NeoCognition is trying to solve.
NeoCognition grew directly out of academic research. Yu Su, the chief executive and a co-founder, is an associate professor at Ohio State University, where he co-directs the OSU NLP group and runs what the company describes as one of the most established AI agent labs in the United States. [1][6] He was named a 2025 Alfred P. Sloan Research Fellow and also received an NSF CAREER Award. [3][6] His co-founders, Xiang Deng and Yu Gu, came out of the same lab; Gu completed his PhD under Su in 2025. [1][6]
Su's group began building large language model based agents before the launch of ChatGPT in late 2022, which placed it unusually early in a field that only became crowded afterward. [1][4] The lab is best known for several research projects that are now widely referenced. Mind2Web, released in 2023, was among the first benchmarks for generalist web agents, with 2,350 tasks drawn from 137 websites across 31 domains. [4][7] SeeAct, released in January 2024, built a web agent on top of GPT-4V and could complete up to half of tasks on live websites under favorable conditions. [4][8] MMMU is a widely used benchmark for multimodal understanding. [1][4] According to the company, work from Su's lab is used across frontier models from OpenAI, Anthropic, and Google. [1][2]
Su had resisted commercializing the research for some time. He has said he turned down earlier pressure from investors before deciding to start the company in 2025, once he became convinced that advances in foundation models made genuinely self-improving, specialized agents feasible rather than aspirational. [4] The founding team brings a combined research background spanning the components of an agent system, including perception, memory, planning, evaluation, and safety. [3]
The technical thesis behind NeoCognition is that the defining feature of human intelligence is not raw knowledge but the ability to keep learning and to specialize. As one summary of the company's positioning put it, "the true power of human intelligence is the ability to continuously learn and specialise." [2] NeoCognition wants agents that build a "world model of work," a structured internal representation of a particular job, its tools, its rules, and its failure modes, and then keep updating that model as they encounter new cases. [1][3]
This draws on Su's academic research interests, which include planning and world models along with memory and non-parametric continual learning. [6] The bet is that an agent which can learn after deployment, rather than being frozen at training time, can reach expert-level reliability in a narrow domain while remaining general enough to be pointed at many different domains. The company says this also makes its agents safer for high-stakes applications, since an agent that understands the constraints of its environment is less likely to take a damaging action. [1][3]
NeoCognition is selling primarily to enterprises and to established software as a service companies, which can use the agents either as autonomous workers or to add agentic features to products they already sell. [1][4] Vista Equity Partners, one of the backers, was reportedly valued in part for the access its large portfolio of enterprise software companies could provide as potential customers. [4] At launch the company employed roughly 15 people, most of them holding PhDs, and said it would use the seed funding to expand research, hire, and move from academic prototypes toward production systems. [2][4]
The $40 million seed round was described as oversubscribed and was co-led by Cambium Capital and Walden Catalyst Ventures, with Vista Equity Partners participating. [1][3] Landon Downs, a managing partner at Cambium Capital, said the firm had "strong conviction in the team's expertise" and believed the research was "charting a new path toward specialised intelligence." [2]
Beyond the institutional leads, the round drew a notable list of angel investors and advisors. Intel chief executive Lip-Bu Tan and Databricks co-founder Ion Stoica both took part. [1][3] Tan praised the founders' research breadth, saying the team had "already developed research that spans every piece of the agent puzzle, ranging from perception to memory, planning, evaluation, and safety." [3] Several prominent academics also invested or signed on as advisors, including Dawn Song, Ruslan Salakhutdinov, and Luke Zettlemoyer, alongside investment vehicles reported to include A&E Investments, Salience Capital Partners, Nepenthe Capital, and Frontiers Capital. [3]
| Funding detail | Reported figure |
|---|---|
| Round type | Seed (oversubscribed) [1][3] |
| Amount raised | $40 million [1][2] |
| Announcement date | April 21, 2026 [1][3] |
| Co-lead investors | Cambium Capital, Walden Catalyst Ventures [1][3] |
| Participating investor | Vista Equity Partners [1][3] |
| Notable angels and advisors | Lip-Bu Tan (Intel), Ion Stoica (Databricks), Dawn Song, Ruslan Salakhutdinov, Luke Zettlemoyer [1][3] |
| Headquarters | Palo Alto, California [1][3] |
| Team size at launch | About 15 [2][4] |
NeoCognition launched into a busy market for agent startups. It competes, broadly, with two groups: companies building general-purpose agents, such as OpenAI, Anthropic, and Google, and companies building more specialized or vertical agents. [2] Its differentiator is the claim that a single architecture can specialize itself through continual learning rather than relying on per-vertical engineering or on a model that stops learning after training. [2][4]
Whether that thesis holds up at scale remains an open question. The company emerged from stealth with research credentials and a funded plan, but with a small team and, at launch, no widely benchmarked commercial product. Self-improving agents that learn safely after deployment are a hard and largely unsolved problem, and the reliability gap Su describes is one the entire field is trying to close. What NeoCognition has in its favor is a founding team that was working on language agents before most of the industry, and a roster of backers betting that the academic head start translates into a real product. The next test is whether the agents can actually learn a job well enough that enterprises stop taking a leap of faith every time they hand one a task.