The Palo Alto startup, spun out of Ohio State University by Yu Su, argues that current agents complete tasks as intended only half the time, a reliability gap it plans to close by giving agents a mechanism to build world models of the domains they operate in, learning on the job as a specialist rather than relying on fixed general training.
NeoCognition, a Palo Alto AI research lab, has emerged from stealth with $40 million in seed funding. The oversubscribed round is co-led by Cambium Capital and Walden Catalyst Ventures, with participation from Vista Equity Partners.
Angel investors and founding advisors include Lip-Bu Tan, CEO of Intel and founding managing partner of co-lead Walden Catalyst Ventures; Ion Stoica, co-founder and executive chairman of Databricks, and AI researchers Dawn Song, Ruslan Salakhutdinov, and Luke Zettlemoyer.
Additional institutional participants include A&E Investments, Salience Capital Partners, Nepenthe Capital, and Frontiers Capital.
The 💜 of EU tech
The latest rumblings from the EU tech scene, a story from our wise ol' founder Boris, and some questionable AI art. It's free, every week, in your inbox. Sign up now!
The company was founded by Yu Su, Xiang Deng, and Yu Gu. Su is a professor at Ohio State University who has led one of the country's most established LLM-based agent research labs since well before the ChatGPT moment, and holds a Sloan Research Fellowship.
He described himself as having initially resisted pressure from venture capital to commercialise his work, until he concluded that advances in foundation models had reached a point where genuinely personalised agents were feasible, at which point he spun the lab out last year.
The problem NeoCognition is trying to solve is reliability. Su's claim, which the company has not independently substantiated in a published benchmark, is that current AI agents successfully complete tasks as intended only around 50% of the time. T
hat figure is broadly consistent with widely-reported findings from AI coding agent evaluations, though specific numbers vary by task type, agent, and evaluation methodology. The consequence, Su argues, is that agents cannot be trusted as independent workers: every task is a gamble.
NeoCognition's response is to give agents a mechanism for rapid specialisation through experience, specifically, by learning to build a “world model” of whatever micro-environment they operate in, capturing its rules, relationships, and constraints through use rather than through pre-training on general data.
The conceptual model draws a direct analogy to human learning. Su's argument is that what makes human intelligence powerful is not its breadth but its plasticity, the capacity to enter a new professional environment and rapidly develop deep domain expertise by internalising how that specific world works.
Current AI agents, optimised for generalism, lack this specialisation mechanism. NeoCognition's thesis is that building it in, as a learnable, autonomous process rather than a manually engineered one, is what separates reliable specialist agents from the current generation of capable-but-inconsistent generalists.
The commercial strategy is primarily enterprise, focused on established SaaS companies rather than consumer end-users.
The pitch to a software vendor is that NeoCognition's agent system can be embedded to create AI workers that improve over time within that vendor's specific operational context, or to power agentic upgrades to existing product offerings.
Vista Equity Partners' participation is framed as a distribution lever: Vista manages one of the largest portfolios of enterprise software companies in private equity, giving NeoCognition potential direct access to software firms actively looking to embed AI at the application layer.
The team has roughly 15 employees, most holding PhDs. The company's specific technical approach has not been disclosed in detail beyond the ‘world model' framing, and no product is yet publicly available.
The $40 million seed is the first institutional capital the company has raised. The timing reflects a broader pattern in AI investment in 2026: capital is increasingly flowing not towards frontier model development, dominated by OpenAI, Anthropic, and a small number of well-capitalised labs, but towards the application and reliability layer, where researchers with agent-specific academic credentials are being recruited and funded at pre-product stage on the strength of their research track record alone.