NeoCognition has raised $40 million in seed funding to build AI agents that teach themselves the rules of a job — a response, founder Yu Su says, to today's agents that succeed roughly half the time and can't be trusted to act unsupervised. The lab grew from Su's research at Ohio State University and was spun into a startup last year; the new capital gives it room to hire researchers, build compute and run longer cycles of experimentation.

What NeoCognition is building NeoCognition describes itself as a research lab focused on creating self‑learning agents that can rapidly master the rules of a given task or profession. The project grew out of work led by Yu Su, an Ohio State University professor who resisted early pitches from venture capitalists before spinning his research into a startup last year. Su says foundational advances in large models have made it possible to rethink how agents learn and personalise. Instead of handcrafted systems tuned for a single industry, NeoCognition wants agents that start general and then teach themselves the micro‑world of a target domain. Su argues humans gain power not from being broad but from our ability to specialise quickly when we join a new workplace or discipline. If agents can build their own internal model of specific environments, he says, they could become reliable autonomous workers rather than tools that sometimes succeed and sometimes fail. Why current agents fall short Su points to inconsistency as the core problem. Today's agents, he says, succeed roughly half the time in completing tasks as intended. That level of unreliability prevents firms from trusting agents to act without supervision. He singled out several recent systems — including Claude Code, OpenClaw and Perplexity’s computer tools — as examples of capable but uneven performers. Part of the trouble is that most agent deployments are engineered for narrow, vertical tasks: they perform well inside the specific box they were built for but struggle when moved even slightly. NeoCognition's approach seeks a middle path: start with a general learner and let it autonomously construct a "world model" for each target job so it can adapt to new rules and relationships without human retooling. Funding, founders and backers The newly disclosed seed round totalled $40 million. The financing was co‑led by Cambium Capital and Walden Catalyst Ventures. Vista Equity Partners also took part, and the round attracted individual backers including Intel chief executive Lip‑Bu Tan and Databricks co‑founder Ion Stoica. Su, who leads the lab and serves as CEO, described the fundraising as a bet by investors that agents can be made more dependable by changing how they learn. The mix of venture firms and prominent tech figures signals investor appetite for projects that promise both research depth and practical results. For a seed round, $40 million is large; it gives NeoCognition room to hire researchers, build compute systems, and run the long cycles of experimentation that agent research requires. How the lab plans to teach agents NeoCognition plans to let agents learn by operating inside simulated or constrained "micro‑worlds" that mimic the real environments they will face. By interacting, observing consequences, and refining internal representations, agents would develop repeated patterns and heuristics akin to human specialists. Su describes the process as building a model of any given micro‑world through autonomous experience. This iterative, experience‑driven learning is presented as different from current, narrowly engineered agent deployments by emphasising autonomous model‑building inside micro‑worlds so agents can generalise and specialise without extensive human retooling. Why this matters At $40 million, the seed is unusually large for the stage and signals clear investor appetite for startups that combine deep research with practical ambitions. That backing gives NeoCognition the resources to scale hiring, buy compute and run the longer experiments the team says are required to prove more dependable agent behaviour.

Related Articles

The seed round was co‑led by Cambium Capital and Walden Catalyst Ventures, with participation from Vista Equity Partners and angels Lip‑Bu Tan and Ion Stoica.

This article was created with AI assistance.