Tuesday brought one of the largest seed rounds in AI agent research this year: NeoCognition, a stealth-mode startup founded by Ohio State University professor Yu Su, emerged publicly with $40 million in funding aimed at solving what Su calls the fundamental flaw in every AI agent on the market today. The round was co-led by Cambium Capital and Walden Catalyst Ventures, with Vista Equity Partners and a handful of prominent angels — including Intel CEO Lip-Bu Tan and Databricks co-founder Ion Stoica — also participating.
A coin-flip reliability rate that enterprises cannot afford
Su's diagnosis of the current agent landscape is blunt. Whether the system in question comes from Anthropic's Claude Code, OpenClaw, or Perplexity's computer-use tools, today's AI agents complete tasks as intended roughly 50% of the time, he told TechCrunch. That number, if accurate, puts the entire category in a curious position: the industry is selling the promise of autonomous digital workers while quietly acknowledging that every deployment is a gamble. For enterprises weighing whether to integrate agents into production workflows, a coin-flip success rate is not a calculated risk — it is a dealbreaker.
"Today's agents are generalists," Su said. "Every time you ask them to do a task, you take a leap of faith." The consistency problem, in his view, goes deeper than model quality or prompt engineering. It stems from the fact that current agents carry no persistent understanding of the specific environment they are operating in — no accumulated knowledge of its rules, edge cases, or institutional logic.
The specialization model: how NeoCognition's approach differs
Su's proposed fix draws directly from how human expertise develops. People are not universally brilliant; what distinguishes a skilled professional is the ability to enter an unfamiliar domain and, over time, build a precise internal model of how it works — its hierarchies, its failure modes, its unspoken conventions. NeoCognition is engineering agents to do exactly that, autonomously and without needing to be custom-built for each vertical from scratch.
"For humans, our continued learning process is essentially the process of building a world model for any profession, any environment," Su explained. "We believe for agents to become experts, they need to learn autonomously to build a model of any given micro world." The distinction from existing approaches matters commercially. Training an agent for autonomous operation in a specific domain is already possible, but it requires bespoke engineering for every use case. NeoCognition's bet is that a single generalist system capable of genuine self-specialization is both more scalable and more durable.
Su spent years leading an AI agent lab at Ohio State before venture capital pressure pushed him toward commercialization — pressure he initially resisted. He changed course last year, after concluding that advances in foundational models had crossed a threshold that made truly personalized, self-improving agents technically achievable rather than aspirational.
Vista Equity as a distribution channel, not just a check
The investor lineup tells its own story. Vista Equity Partners, one of the largest private equity firms focused on enterprise software, brings something beyond capital: a portfolio of established SaaS companies actively looking to embed AI capabilities into their products. Su flagged this explicitly, describing Vista's involvement as providing direct access to a roster of potential enterprise customers — the precise market NeoCognition is targeting. For a 15-person company, the majority of whose staff hold PhDs, that kind of warm commercial pipeline is worth considerably more than the dollar figure on the term sheet.
The angel roster reinforces the enterprise-first signal. Lip-Bu Tan, now running Intel, has spent decades at the intersection of semiconductor infrastructure and enterprise software. Ion Stoica's work at Databricks sits at the core of how large organizations manage and process the data that AI systems depend on. Neither is a typical consumer-tech angel.
NeoCognition plans to sell primarily to established enterprises and SaaS platforms, positioning its agent systems as either standalone worker products or as infrastructure that existing software vendors can layer into their offerings. With the seed round closed and the team now operating in the open, the next pressure point is a working product that can demonstrate, in a real enterprise environment, the reliability gains Su is promising. The 50% benchmark he cited for competitors sets the bar — and the expectation — at the same time.
