NeoCognition raises $40M to fix the part of AI agents everyone keeps hand-waving away

NeoCognition, a startup spun out of Ohio State professor Yu Su’s AI agent lab, has emerged from stealth with a $40 million seed round led by Cambium Capital and Walden Catalyst Ventures. Vista Equity Partners joined, along with angels including Intel CEO Lip-Bu Tan and Databricks co-founder Ion Stoica.

The funding matters. So does the thesis. NeoCognition wants to build self-learning AI agents that become reliable specialists in whatever domain they’re dropped into.

That’s a good place to aim, because it gets at the weakest part of the current agent push.

There’s no shortage of demos showing agents clicking through browsers, writing code, filing support tickets, and chaining tools together. What’s still missing is consistent execution. Su told TechCrunch that current agents complete tasks as intended only about 50% of the time. The exact number depends on the benchmark, the workflow, and how generously you define success, but the underlying point holds up. Reliability still collapses when the environment gets messy, state changes unexpectedly, or the task calls for domain judgment.

For technical teams, that’s the line between a lab toy and production software.

The bet: agents need a world model for the job

Su’s framing is straightforward. Humans are general learners, but they become useful through specialization. A new finance analyst, support rep, or operations manager doesn’t just memorize steps. They build an internal model of how that small system works: which rules are rigid, which exceptions come up all the time, who approves what, what breaks downstream, and which signals actually matter.

NeoCognition wants agents to do the same. In Su’s words, the goal is for agents to autonomously build a model of a “micro world” and learn to operate inside it like an expert.

That gets closer to the real problem than the usual talk about autonomous coworkers. Most enterprise work doesn’t fail because a model lacks general knowledge. It fails because the model doesn’t understand the local environment. It doesn’t know the odd CRM conventions, the undocumented exception paths, the internal policy logic, or the consequences of taking the wrong action in the wrong order.

Foundation models are broad. Enterprise work is narrow and fussy.

The companies that matter in this market will be the ones that close that gap.

Why this is hard

Plenty of agent startups already say their systems learn over time. Usually that means some mix of prompt tuning, retrieval, workflow memory, reinforcement from human feedback, or storing successful trajectories and replaying them.

Useful, yes. Human-like learning, no.

If NeoCognition is serious about self-learning specialization, it needs more than cached examples and a larger context window. It needs a way for an agent to build structured internal representations of a domain, update those representations as the environment changes, and use them to plan actions with lower error rates over time.

That raises a few obvious questions.

What is the “world model”?

Is it symbolic, latent, retrieval-backed, or some hybrid?

A serious agent system for enterprise work probably needs all of the above:

• a latent model for general pattern recognition
• explicit symbolic state for workflows, policies, and constraints
• retrieval over documents, logs, tickets, and prior actions
• persistent memory tied to tasks, users, and systems

A single giant model won’t absorb all of that cleanly if you care about auditability, debuggability, or cost.

How does it learn safely?

Self-learning sounds great until the agent learns the wrong lesson from noisy outcomes. In real systems, success signals are often ambiguous. A support ticket closed quickly may still be mishandled. A code change may pass tests and still introduce a security flaw. A procurement agent might optimize for speed and quietly violate policy.

If NeoCognition wants enterprise buyers, it needs guardrails around adaptation. That means permissioning, policy checks, reversible actions, human review loops for higher-risk decisions, and a clear separation between learning the domain and changing live behavior without approval.

What does specialization cost?

The economics matter. A self-improving expert agent sounds appealing until every deployment turns into its own training pipeline with heavy inference costs, long warm-up periods, and brittle evaluation.

This is where a lot of agent ambition runs into enterprise procurement reality. Buyers don’t want a research project. They want something that reaches acceptable accuracy quickly, stays within budget, and doesn’t need an internal ML team to babysit it.

NeoCognition hasn’t said much publicly about implementation, so for now the technical case is still mostly thesis.

NeoCognition is aiming at the right bottleneck

The market has spent the past two years rewarding flashy agent interfaces and spending too little time on failure rates.

That imbalance was always odd. UI matters. Tool orchestration matters. Browser automation matters. None of that fixes the central production problem. Agents are still inconsistent in dynamic environments.

That shows up clearly in coding tools.

Claude Code, OpenClaw, and Perplexity’s computer-use style products can be genuinely useful, but anyone who uses them seriously knows the pattern. You get a run of impressive output, then a baffling miss on something routine. The issue usually isn’t raw intelligence. It’s incomplete situational understanding. The model doesn’t fully grasp the local codebase, the hidden constraints, or the practical meaning of done.

Su’s diagnosis lands because it matches that experience.

An agent that can build real competence inside a bounded environment would be worth far more than another system that demos well on a clean benchmark.

Why Vista’s participation matters

Vista Equity Partners joining the round suggests something about the go-to-market plan.

NeoCognition plans to sell agent systems to enterprises and established SaaS vendors, which makes sense. Those buyers already own narrow domains full of repetitive, rules-heavy work. They also have the data needed to teach a system how those domains behave.

That could be a strong wedge. If you’re selling into a portfolio of software companies trying to add AI features without shipping unreliable agents into customer workflows, self-learning specialist agents is at least a serious pitch.

It also raises the bar. Enterprise software customers won’t tolerate fuzzy claims for long. If NeoCognition gets distribution through software vendors, it will need to show measurable gains in task completion, error reduction, adaptation speed, and cost per workflow.

Nobody buying infrastructure-grade AI wants a lecture about human learning. They want dashboards, evals, rollback controls, and security docs.

What developers and AI teams should watch

For engineers, the interesting question isn’t whether NeoCognition can make agents “learn like humans.” That phrase can mean almost anything.

The useful questions are narrower.

1. How is memory handled?

Does the system maintain long-term, structured memory across tasks, or is it mostly summarizing prior interactions into prompt context? Memory is where a lot of agent products quietly break.

2. What evals show specialization?

A real specialist agent should improve with repeated exposure to the same environment. That should show up in metrics like:

• task success rate over time
• reduced need for human correction
• lower token and tool-call cost per completed task
• fewer policy violations or invalid actions
• faster recovery from edge cases

If the startup can’t show those curves, the pitch is still aspirational.

3. Can it operate under constraints?

Enterprise environments are permissioned, fragmented, and full of ugly APIs. A good agent has to work with partial visibility, stale data, approval gates, and systems it can’t directly control.

4. What’s the security model?

Self-learning systems raise obvious questions around data retention, cross-customer isolation, prompt injection, malicious tool outputs, and poisoning through bad feedback signals. If the product learns from enterprise behavior, buyers will want clear answers on where that learning is stored and how it’s separated.

5. How model-dependent is it?

If NeoCognition’s results depend heavily on whichever frontier model is hottest this quarter, that limits defensibility. If it has a strong orchestration, memory, planning, and evaluation layer that survives model churn, that’s more interesting.

A $40 million seed says investors think the window is open

That’s a big seed, though not an unusual one by 2026 standards, especially for a research-heavy AI company with a team of around 15 people, most of them PhDs. Investors are still willing to spend heavily on technical founders going after foundational agent problems.

And yes, this is a better use of AI capital than another thin wrapper promising “AI employees” without any real answer on reliability.

Still, there’s a gap between naming the right problem and solving it. Plenty of labs have argued that agents need memory, planning, self-reflection, better environment models, or domain adaptation. Most run into the same wall: keeping the system robust while it learns in the wild.

NeoCognition deserves attention because it’s going after the part of agent systems that actually blocks adoption. The interface matters less. The demo matters less. The hard part is getting an agent dependable enough that a company will let it do useful work without a human shadowing every step.

That’s the standard. If NeoCognition can hit it, this seed round will look cheap. If not, it’ll end up on the same list as a lot of agent startups that diagnosed the problem correctly and still failed to ship the product.