NeoCognition's investment of $40 million in self-learning AI agents

      The Palo Alto startup NeoCognition, founded by Yu Su and originating from Ohio State University, claims that current AI agents only successfully complete tasks as intended about half the time. To address this reliability issue, the company aims to enable agents to create world models of their operating domains, allowing them to learn as specialists while working rather than depending on fixed general training.

      NeoCognition has recently emerged from stealth mode with $40 million in seed funding. The round, which was oversubscribed, is co-led by Cambium Capital and Walden Catalyst Ventures, with additional participation from Vista Equity Partners.

      Among the angel investors and founding advisors are Lip-Bu Tan, CEO of Intel and founding managing partner at Walden Catalyst Ventures; Ion Stoica, co-founder and executive chairman of Databricks; along with AI researchers Dawn Song, Ruslan Salakhutdinov, and Luke Zettlemoyer. Further institutional participants include A&E Investments, Salience Capital Partners, Nepenthe Capital, and Frontiers Capital.

      The company was co-founded by Yu Su, Xiang Deng, and Yu Gu. Su, a professor at Ohio State University, has led a prominent LLM-based agent research lab long before the rise of ChatGPT and holds a Sloan Research Fellowship. Initially, he resisted pressure from venture capital to commercialize his work until he realized that progress in foundational models had made genuinely personalized agents attainable, prompting him to spin off the lab last year.

      NeoCognition aims to tackle the issue of reliability. Su has stated, though this has not been independently verified in a published benchmark, that current AI agents only succeed in completing tasks approximately 50% of the time. This figure aligns with previously published findings on AI coding agent evaluations, although specific metrics may differ based on task type, agent, and evaluation methods. Su argues that this unreliability means agents cannot be viewed as dependable independent workers, as every task is uncertain.

      To counter this, NeoCognition proposes that agents be equipped with a mechanism for rapid specialization through experience, specifically by learning to develop a "world model" of the micro-environment they inhabit, identifying its rules, relationships, and constraints through practical use instead of pre-training on generalized data.

      This conceptual model directly parallels human learning. Su posits that human intelligence is powerful not due to its breadth but its adaptability—the ability to enter a new professional setting and quickly gain deep domain expertise by understanding how that specific world functions. Current AI agents, designed for generalism, lack this specialization feature. NeoCognition contends that incorporating this as a learnable, autonomous process will distinguish reliable specialist agents from the current generation of capable but inconsistent generalists.

      The company's commercial strategy predominantly targets enterprises, focusing on established SaaS companies rather than individual consumers. NeoCognition pitches its agent system to software vendors, suggesting it can be integrated to create AI workers that improve within the specific operational context of the vendor or enhance existing product offerings.

      Vista Equity Partners’ involvement serves as a distribution strategy: they manage one of the largest portfolios of enterprise software companies in private equity, potentially providing NeoCognition direct access to software firms eager to integrate AI into their applications.

      The NeoCognition team consists of around 15 employees, most of whom hold PhDs. The specific technical approach has not been detailed beyond the framework of the 'world model,' and no product is currently available to the public. The $40 million in seed funding marks the company's first raised institutional capital. This timing reflects a broader trend in AI investment in 2026, where funding is increasingly directed not at pioneering model development—led by OpenAI, Anthropic, and a limited number of well-funded labs—but toward application and reliability. This shift has resulted in recruitment and funding for researchers with agent-specific academic expertise at pre-product stages based solely on their research achievements.