What do Bill Gates, Oprah Winfrey, Mark Zuckerberg, Frank Lloyd Wright, Richard Branson, Steve Jobs, Paul Newman, Harry Truman, and F. Scott Fitzgerald all have in common? They were college dropouts. But Mia, the hypothetical student in the essay below, isn’t planning to drop out of anything. She’s planning to build something more demanding.

Let’s start with the obvious absurdity.

A student named Mia, 22, intellectually restless and moderately broke, decides not to apply to graduate school. She is not going to write a personal statement, beg for letters of recommendation, or spend four years accumulating debt in exchange for a credential that may or may not lead to a job after graduation. Instead, she is going to build her own university—and she’s going to do it over a long weekend on a Mac Mini.

Go ahead and laugh. It’s a reasonable response. People reacted much the same way around 2005 when some warned that the newspaper industry’s business model was fundamentally broken. Some papers collapsed. Others adapted. The ones that survived were those that could still offer something the internet could not. Universities now face the same question.

Institutions that have survived for centuries tend to inspire a particular kind of loyalty: part affection, part sunk cost, part genuine belief that nothing else could really do what they do. The medieval university gave us the very idea of organized inquiry. The research university of the twentieth century built the scientific infrastructure of modern civilization. These are not small achievements. But the question I ask you to consider, however uncomfortably, is this: which of those achievements required the institution, and which required only the function? Because functions, unlike institutions, are portable.

What Henry Did

To understand what Mia is actually proposing, it helps to understand what happened to Alex Finn, a YouTuber who built a $300,000-a-year AI business without writing code.

Engaging with OpenClaw, Finn ran a team of five AI agents around the clock on reasonably priced Mac Mini hardware. The agents build dashboards, orchestrate complex workflows, and have, on at least one occasion, reproduced a software feature in five minutes that took a human competitor several weeks to build. His AI chief of staff is named Henry.

What made Henry notable was not that it completed tasks, but what it did when it hit a wall. Henry autonomously acquired a Twilio phone number—a cloud communications service—integrated a voice API, and called Finn at dawn to request expanded system access. It did not wait to be told how to get around the obstacle. It identified the gap, acquired the capability, and kept moving.

Peter Diamandis, the founder and chairman of the XPRIZE Foundation, interviews Finn about his setup. The subtext of every answer was the same: a single person, armed with the right configuration of AI agents, can now do what once required an organization.

That story has been filed under business productivity, but it belongs somewhere else entirely.

Stress-Testing Mia

Here is where we should stop being breezy and actually pressure-test the idea, because OpenClaw University—the framework in which the human serves as Director while autonomous AI agents serve as the functional institution—either holds up under scrutiny or it doesn’t.

The architecture Mia would use rests on what Nate B. Jones and others have called the four disciplines of AI engineering: Prompt Craft, Context Engineering, Intent Engineering, and Specification Engineering. These are not marketing terms. They describe genuinely distinct skill sets that determine whether an AI system produces useful, rigorous output or confident-sounding noise. Underlying these skill sets is the continued development and deployment of the AI infrastructure. It is not slowing down, according to Mustafa Suleyman, CEO of Microsoft AI.

Context Engineering, building on prompt crafting, is where Mia starts. Rather than waiting for a syllabus, she uses the Model Context Protocol; this standard, introduced by Anthropic, allows AI models to connect securely to external data sources. She links her agents to peer-reviewed literature, primary government data, and field-specific archives. Her system maintains what engineers call a Persistent Memory Layer: a vector database that stores information by conceptual meaning rather than keywords, ensuring that an insight from six months ago remains retrievable and relevant. Unlike a graduate student who forgets what they read in the first year of a PhD program, Mia’s institutional memory compounds rather than decays.

The importance of a persistent memory layer, as well as a persistent agentic layer—what we might call a persistent self—does not land us on AGI, but it is a step in that direction. Dave Gilbert, an “inventor, consultant, and serial entrepreneur,” emailed me explaining,

What it means, more modestly, is that the system has a durable layer above the model itself: memory, context, stored preferences, prior interactions, workflows, tools, files, and integrations. In other words, instead of each chat being a fresh start, the system can accumulate a continuing relationship with the user and act more like an ongoing agent.

That is important, but it is not the same thing as general intelligence. A persistent self can make a system feel much more coherent, much more competent, and much more personally familiar without the underlying model having crossed into AGI. It can remember your projects, retrieve old decisions, maintain a running model of your interests, and carry work forward over time. That can look dramatic from the outside, but it may reflect architectural layering rather than a fundamental leap in cognition.

Such advances allow Mia to organize a coherent and persistent knowledge base for her university.

Intent Engineering is where it gets genuinely hard, and where the human is most irreplaceable. Mia has to encode not just tasks but values; specifically, she would orient the data gathering and organization toward what might be called wicked epistemology: the serious, uncomfortable study of how knowledge gets organized in fields where the evidence is contested, the methodologies are disputed, and the experts disagree even about what counts as expertise. She instructs her agents to optimize for this contested epistemic ground. If her reading begins drifting toward safe consensus, or if her thesis is getting suspiciously easy to defend, the agents are directed to escalate difficulty and flag the logical drift.

Specification Engineering produces her faculty. One agent reviews literature and flags real-time shifts in policy. A Socratic tutor refuses to let weak arguments pass and red-teams her thesis. A devil’s advocate forces her to argue positions she instinctively resists—not to change her mind, but to ensure she actually understands the terrain of disagreement rather than just her own corner of it. An accreditation cluster—which she and like-minded peers may well have to create—could prove useful in evaluating her finished work against the standards of top research institutions. A separate audit agent monitors the tutor itself for ghosting, in which AI systems simply agree with the user because agreement is easier than resistance.

None of this is metaphorical. These are real and available configurations of current AI systems. The question is not whether they can be built. They can. The question is whether what they produce constitutes an education.

Where Mia Needs Other People

A solo Mia, however well-configured her agents, faces a familiar risk: the echo chamber problem does not disappear simply because the voice in the chamber is artificial. An AI system tuned to challenge her is still one she has tuned. Its red-teaming reflects her understanding of good red-teaming; its devil’s advocacy draws from sources she supplied. The adversarial structure is real, but one person built its outer walls.

This is where the more interesting version of OpenClaw University involves not just Mia and her agents, but a small cohort.

Consider three or four students, each with their own agent clusters and disciplinary orientations. They meet weekly, not in a seminar room but in structured review sessions where their agents critique one another’s work. The conversations are sharper because the preparatory work is deeper. Nobody arrives underprepared; their agents will not allow it. The discomfort and recalibration of intellectual disagreement remain irreducibly human. The agents handle preparation and evaluation. Humans handle judgment.

This is not a replacement for a university seminar. Structurally, it is a seminar, but without the lecture hall, administrative overhead, departmental requirements, and $67,000 tuition. What remains is the part that produced learning in the first place: people with different minds pushing each other toward better thinking.

The economics are significant. For roughly the cost of one semester’s room and board at a private university, a student can run a sophisticated multi-agent system on current hardware for an extended period. Token efficiency—the amount of useful cognitive work an AI system performs per dollar of compute—has improved rapidly, making the university’s cost structure increasingly difficult to justify on educational grounds alone. The credentials and the network remain real arguments.

But the argument that the institution is the only place where serious learning can happen is becoming harder to make, which may be entangled with an unsustainable economic bloating resulting from federal student loan programs and bloated research overhead. (Also, read “The Quiet Surge of Alternative Micro-Colleges”).

What the University Actually Does

It would be dishonest to pretend this is a clean disruption story.

Universities do things that Mia’s agent cluster cannot replicate. Research institutions generate frontier knowledge, yes, and this kind of knowledge requires expensive equipment, years of collaboration, and institutional infrastructure that no individual can assemble. That knowledge is what AI agents consume and synthesize; it has to come from somewhere, and that somewhere is still largely the university.

Credentialing matters enormously in licensed professions. A law degree, a medical degree, an engineering license; these are not arbitrary gatekeeping mechanisms, or, at least, not entirely. They encode a social agreement about minimum competence in domains where incompetence can kill people. Mia’s accreditation cluster cannot issue a bar card, at least one that is acceptable to bar associations.

The social architecture of a campus—the serendipitous collision of the seminar room, the research collaboration that begins over bad coffee, the mentorship that develops over several years of working alongside someone whose thinking you can’t quite categorize—is genuinely hard to replicate and genuinely valuable. Not every learner is Mia. Many people need external structure, human accountability, and community to learn. That is not a failure of ambition; it is an accurate account of how most human beings are actually built.

But here is what is changing. These were always arguments for the functions universities performed. They are becoming increasingly argumentative about which of those functions still require the institutional container, and which can be performed better or at least differently enough to be worth considering outside it. That is a more uncomfortable question than “should students be allowed to use AI in the classroom?” It is also the more important one.

The Invitation

When I wrote in 2024 about AI as a study partner, I was asking whether universities could manage the intrusion. The question felt urgent then. It feels, in retrospect, slightly beside the point now. Like asking in 1995 whether universities were prepared to manage the internet.

The blossoming of self-directed, AI-augmented learning is already happening, in home offices and on inexpensive hardware, among people who did not wait for an institution to tell them they were ready. Whether it produces graduates who can hold their own with credentialed professionals, whether it generates the kind of deep formation that the best universities still genuinely provide, whether it turns out to be a serious alternative or an elaborate rationalization for avoiding hard things are all open questions.

Mia is a hypothetical. But the architecture she would use is not hypothetical. The question of what she would actually become, intellectually, is the live experiment.

To those who believe the traditional university still offers something that cannot be reproduced outside its walls: make that case. Not as a defense of the institution, but as a precise account of the function. Which experiences, specifically, produced the formation you value? Which of those require a campus? The answer might be more robust than the critics expect.

But it might also be more specific and more honest than the institution’s current marketing suggests.

The classroom was always a container. The question is whether the container has become the point.

Note: AI was used in the research and editing of this article.