AI startup funding is moving at full speed, and investors are increasingly keeping their eyes on researchers who can turn academic labs into real products. The race to find the next big breakthrough in AI shows no signs of slowing down, and the checks keep getting bigger, even for companies that haven’t launched a public product yet.
But there’s one detail that still bothers a lot of people in the industry: the AI agents available today only get things right about half the time.
Fifty percent reliability.
That’s way too low for any company to seriously bet on this technology. Whether you’re using Claude Code, OpenClaw, or Perplexity’s computer tools, the current landscape is the same: every time you ask an agent to perform a task, you’re basically flipping a coin.
That’s exactly the gap NeoCognition wants to close.
The startup just emerged from stealth mode with a $40 million seed round in the bank and a pitch that turns heads: building agents capable of self-learning, meaning they learn and specialize on their own, the same way a human does when stepping into a new environment or profession. 🚀
If this actually works in practice, it could be a game-changer for the future of autonomous agents.
Where NeoCognition comes from and who’s behind it
NeoCognition is described by its founder as a research lab developing self-learning AI agents. Leading the project is Yu Su, a professor at Ohio State University who runs a research lab focused on AI agents. Su shared that he initially resisted pressure from VCs to commercialize his academic work. He held out for quite a while, until he finally decided to make the leap last year, when he realized that advances in foundation models could make truly personalized agents possible.
That decision to leave academia wasn’t made on impulse. Su saw a technical window of opportunity that didn’t exist before: the evolution of large language models had reached a point where the technological foundation needed to build agents that learn continuously had finally become viable. And when a researcher of that caliber decides to start a company, the market pays attention.
Receive the best innovation content in your email.
All the news, tips, trends, and resources you're looking for, delivered to your inbox.
By subscribing to the newsletter, you agree to receive communications from Método Viral. We are committed to always protecting and respecting your privacy.
The $40 million in funding was co-led by Cambium Capital and Walden Catalyst Ventures, with participation from Vista Equity Partners and heavyweight angel investors, including Lip-Bu Tan, CEO of Intel, and Ion Stoica, co-founder of Databricks. The fact that the company raised this amount while still in seed stage, with no public product and operating in stealth mode, says a lot about the confidence the industry is placing in the core idea behind NeoCognition and in its founder’s track record.
Vista Equity Partners’ involvement, in particular, carries important strategic significance. As one of the largest private equity firms in the software space, Vista can give NeoCognition direct access to a vast portfolio of companies looking to modernize their products with artificial intelligence. This isn’t just money in the bank — it’s a built-in distribution channel from day zero.
The real problem with AI agents today
To understand what NeoCognition is trying to solve, it’s important to put the current state of AI agents into context. An agent, in the technical vocabulary of the field, is a system based on artificial intelligence that receives a goal and takes a sequence of decisions and actions to achieve it, without needing a human to guide every step of the process. Sounds simple in theory, but in practice the agents available today face a very concrete limitation: they’re unreliable, make mistakes on chained tasks, and struggle with contexts that weren’t well represented in the original training data.
As Yu Su put it pretty bluntly: today’s agents are generalists. Every time you ask them to do a task, you take a leap of faith. The success rate on complex tasks tends to hover around 50%, which makes using these systems too risky for business-critical applications.
That number matters a lot more than it seems. When a company starts thinking about automating processes with AI agents, they’re not thinking about simple tasks like answering a generic email. They’re thinking about entire workflows — contract analysis, advanced technical support, order processing, candidate screening, or operations monitoring. For those use cases, failing half the time isn’t an inconvenience — it’s a serious operational problem that can cost money, time, and even the company’s reputation.
That’s why many organizations still prefer to keep humans at the center of these processes instead of fully trusting the available agents. And that’s precisely why agents haven’t yet become the independent, reliable workers that many promised they would be by now.
NeoCognition’s vision: learning the way humans learn
NeoCognition’s approach starts with an observation about human intelligence itself that is both obvious and profound. Yu Su argues that while human intelligence is broad and generalist, its real power lies in the ability to specialize quickly. When we step into a new environment or profession, we can master its unique rules, relationships, and consequences at an impressive speed.
That’s the dynamic NeoCognition wants to replicate in its agents.
For humans, our continuous learning process is essentially the process of building a world model for any profession, any environment, Su explained. He believes that for agents to become true specialists, they need to autonomously learn how to build a model of whatever micro-world they’re placed in.
The self-learning approach proposed by NeoCognition steps in right here as an attempt to change the equation. The core idea is that, instead of relying on a static model that was trained up to a certain point and then frozen, the company’s agents would be capable of continuing to learn from the interactions, mistakes, and specific contexts of each environment they operate in. This closely mirrors how a human professional develops over time: they arrive with a general set of knowledge and gradually specialize as they face real situations, learn from the results, and adjust their behavior based on what works best in that specific context.
Su sees this capacity for rapid specialization as the critical missing link to make AI work reliably on its own. And that’s where NeoCognition sets itself apart from other approaches on the market. 🤖
What sets NeoCognition apart from other solutions
An important point Su made sure to highlight is that it’s already possible to train agents for autonomous tasks today. The problem is that these agents need to be custom-engineered for a specific vertical. In other words, you build an agent that works well for e-commerce customer support, but it can’t be transferred to financial analysis or supply chain management without significant rework.
NeoCognition is building something different: generalist agents with self-learning capabilities that can specialize in any domain. That’s the fundamental difference. Instead of creating one agent per vertical, the idea is to create an agent that starts as a generalist and becomes a specialist through actual use, without constant manual intervention.
Think of it this way: instead of hiring a consultant who comes pre-programmed to solve only one type of problem, you hire someone extremely smart and adaptable, who can dive into any area and become a go-to expert within weeks. That’s the analogy that best describes what NeoCognition is trying to build.
Self-learning as a technical differentiator
The concept of self-learning isn’t new in artificial intelligence research, but putting it into practice in a robust and scalable way is a considerable technical challenge. Most current systems learn during the training phase and then operate in inference mode — meaning they apply what they’ve learned without modifying their own parameters. Retraining a large model is an expensive, time-consuming process that requires significant infrastructure, making it impractical to do continuously for every agent in production.
What NeoCognition appears to be developing is an architecture that allows the agent to update its behavior more dynamically, without necessarily going through a full retraining cycle every time it encounters new relevant information or experience. The micro-world idea, mentioned by Su, suggests that the agents would build internal representations of the environments they operate in — something like a mental map that becomes more detailed and accurate as the agent accumulates experience in that context.
This approach has some really interesting practical implications. An agent with genuine self-learning capability could, for example, progressively get better at understanding a company’s specific vocabulary, its customers’ behavioral patterns, the exceptions that exist in internal processes, and the undocumented preferences of the people who use the system every day. Over time, that agent would stop being a generic tool and become something much closer to a specialized collaborator — one that deeply understands the context it operates in and can make more precise decisions precisely because of that accumulated contextual experience.
From a technical standpoint, there are still many open questions about exactly how NeoCognition plans to implement all of this. The company hasn’t released deep technical details about its architecture, which is common for early-stage startups that need to protect their intellectual property. What we do know is that the current team has around 15 employees, most of whom hold PhDs, which reinforces the company’s research DNA and its commitment to genuine technical innovation.
The business model and target market
NeoCognition plans to sell its agent systems primarily to enterprises, including established SaaS companies. The intended use model covers two main scenarios: companies that want to build worker agents using NeoCognition’s technology, or companies that want to use those agents to enhance existing product offerings.
This B2B positioning makes a lot of strategic sense. Instead of competing directly with consumer-facing AI personal assistants — where the competition with Google, OpenAI, and Microsoft is brutal — NeoCognition is targeting the enterprise segment, where demand for reliable and specialized agents is massive and the willingness to pay for solutions that actually work is significantly higher.
And this is where Vista Equity Partners’ presence on the cap table comes back into play. With access to Vista’s portfolio, which includes dozens of large-scale software companies, NeoCognition has a potential client pipeline from day one. For a seed-stage startup, having that kind of distribution advantage is rare and extremely valuable.
What to expect from the next steps
With $40 million in hand, NeoCognition has enough capital to expand its research and engineering team, push forward on developing its self-learning technology, and start testing its agents with strategic partners before a broader launch. The typical playbook for startups at this stage involves closing deals with anchor companies willing to test the technology in exchange for early access and influence over the product roadmap — which also generates real-world data that feeds back into development.
The AI agent market is growing fast, with players like Salesforce, Microsoft, Google, and a host of smaller startups all racing to launch solutions in this direction. What sets NeoCognition apart in this landscape isn’t just the technical proposition, but the combination of solid academic credentials, a long-term vision grounded in real research, and the capital to execute patiently and consistently.
In a space where most companies are still trying to solve the basic reliability problem of agents, coming in with a self-learning approach and autonomous world-model building is an ambitious play. But with the right team, adequate funding, and aligned strategic partners, NeoCognition has the ingredients to turn that ambition into something real.
What’s clear, looking at all of this, is that the conversation around AI agents is shifting in tone. It’s moving past the generic hype and entering a more mature phase, where the central question is no longer whether agents will exist, but whether they’ll be good enough to reliably replace or augment human work. NeoCognition is betting that the answer to that question runs through self-learning, and with the team and funding it has now, it’s going to be very interesting to watch whether that bet turns into technology that truly changes the game. 👀
