The archive · AI & Models · Strategic decision · 2025–2026
NeoCognition bets agents must learn on the job; Ohio State prof raises $40M seed
Yu Su spun his Ohio State agent lab into NeoCognition: agents that learn on the job to become experts, backed by a $40M seed.
NeoCognition
What the business is
NeoCognition is a Palo Alto research lab, spun out of Yu Su's Ohio State agent lab, building AI agents that continuously learn the rules and workflows of each environment they operate in and specialize into domain experts, sold primarily to enterprises and SaaS companies.
Starting capital：$40M seed, announced 2026-04-21.
How it started
Yu Su runs one of the most established academic AI-agent labs in the US at Ohio State, where his team began building large-language-model agents before the ChatGPT moment and produced foundational work such as Mind2Web, MMMU and SeeAct. He says he long resisted venture money, then changed his mind in 2025 when foundation-model progress convinced him agents could become genuinely personalized and self-improving, and spun his research out into NeoCognition.
What happened
NeoCognition operates in stealth as a lab with about 15 employees, most holding PhDs. On 2026-04-21 it emerged from stealth with a $40M oversubscribed seed round co-led by Cambium Capital and Walden Catalyst Ventures, with Vista Equity Partners participating and angels and founding advisors including Intel CEO Lip-Bu Tan, Databricks co-founder Ion Stoica and AI researchers Dawn Song, Ruslan Salakhutdinov and Luke Zettlemoyer. Su's pitch to TechCrunch is that today's agents, whether Claude Code, OpenClaw or Perplexity's computer tools, complete tasks correctly only about half the time because they are generalists, whereas NeoCognition's agents build a structured model of each micro-world and specialize on the job, the way a human entering a profession rapidly masters its rules and relationships.
No ending yet — it is still running.
Background
Yu Su, an Ohio State professor and Sloan Research Fellow, spent years running one of the country's most established academic AI-agent labs, contributing foundational work such as Mind2Web, MMMU and SeeAct before spinning it out in 2025 as NeoCognition. He told TechCrunch he initially resisted venture pressure, but decided the leap was worth it once foundation-model progress made genuinely personalized, self-improving agents plausible.
The thesis is that current agents fail because they are generalists: Su says Claude Code, OpenClaw and Perplexity's computer tools complete tasks correctly only about 50% of the time, which keeps them out of trusted, independent work. NeoCognition is building agents that continuously learn the structure, workflows and constraints of the environments they operate in and specialize into domain experts, mirroring how humans master a profession by building a model of their micro-world rather than requiring custom engineering for every vertical.
On 2026-04-21 NeoCognition emerged from stealth with a $40M oversubscribed seed round co-led by Cambium Capital and Walden Catalyst Ventures, with Vista Equity Partners participating and angels and founding advisors including Intel CEO Lip-Bu Tan, Databricks co-founder Ion Stoica and prominent AI researchers. The lab, about 15 people mostly holding PhDs, plans to sell its agent systems primarily to enterprises and established SaaS companies, with Vista's portfolio of software businesses as a natural first channel.
What has to be true
- The problem is concrete and widely felt: enterprise agents that only work half the time cannot be trusted as independent workers.
- Self-learning specialization is a real differentiator against both general-purpose agents and custom-engineered vertical tools.
- A decade of influential Ohio State research gives NeoCognition credibility and a talent pool no seed-stage rival can easily copy.
- Investors brought a channel: Vista Equity's software portfolio gives the lab immediate access to companies modernizing products with AI.
- The open risk is that learning on the job in high-stakes settings is unproven, and safety questions will decide whether enterprises adopt it.
What can be applied
An academic founder can win a giant seed round on research pedigree, but the bet only pays if on-the-job learning makes agents reliable enough for real enterprise work.
Aftermath
As of 2026-09-05 NeoCognition is out of stealth with $40M committed and about 15 mostly-PhD employees still building self-learning expert agents; no product launch, customer names or further financing have been announced.
Sources
- AI research lab NeoCognition lands $40M seed to build agents that learn like humans
- NeoCognition Emerges from Stealth With $40 Million Seed Round to Advance Specialized Intelligence and Expert Agents
spotted an error? The archive wants to know.
Your turn
You just read one. Describe what you are building, and see who is betting on the same thing.
Free account · 3 free questions · no card