I'm an AI professional rambler. 🥲 I was about to have a drink early in the evening and go to bed, but suddenly I'm rambling about random thoughts again (!). For reference, I graduated from a mediocre undergraduate program and a Japanese mediocre (what we call an Imperial) university graduate school. Of course, I wasn't good at studying and didn't want to study, but I'm one of those lucky cases you often see around who didn't fail academically. 😁 When I listen to stories from my seniors, juniors, and colleagues around me, I really don't understand why things like this are the same across East and West ㅎㅎ As always, please be warned that this is a long post 🙏

I had a strange sense of déjà vu when I had multiple AI agents running and working on tasks.

One writes code, one reviews it, one organizes documentation, one finds counterexamples, one analyzes the causes of test failures. I look at the results and give new instructions.

There's insufficient basis for that. Well, I'll ask another agent about it. Review this part again. The draft is fine, but the structure is weak. Run the tests and bring me an analysis of why it failed.

At first, I thought of it as just a development workflow for the AI era. But the more I thought about it, the more familiar it felt. Isn't this a graduate school lab? Where a professor divides work among PhD and master's students, asks them to find prior research, run experiments, write paper drafts, picks them apart at lab meetings, and tells them to do it again.

It was a joke I brought up to laugh about, but thinking more about it, it didn't feel funny. Because beneath the surface of this cutting-edge technology called AI orchestration, I could see the old laboratory structure reflected exactly as it was.

First, let me draw a line

This is not professor-bashing. There are definitely advisors who treat students as fellow researchers and help with their intellectual growth. Such people exist now and will continue to exist.

But the existence of good advisors doesn't erase the repeated power asymmetry within the graduate school system. A good research lab can be created when a good person holds that position, but when someone who hasn't sufficiently reflected on their power wields excessive authority, a lab easily becomes a small kingdom. The problem is the structure itself where results depend on individual character. If the only way to avoid exploitation is to encounter a good person, then the structure is broken.

So this essay is not aimed at the profession of professor, but at the way power operates through that title.

Did graduate school teach knowledge or the use of power?

Graduate school is a space for knowledge production. You read papers, define problems, conduct experiments, analyze data, and create new knowledge. In the best sense, it can be the most intellectually dense place to learn.

But graduate school in reality is also a labor organization. Project reports, experiments, coding, data organization, paper drafts, conference preparation, lab administration, mentoring junior students, and sometimes even the professor's odd jobs. All this work is bundled under the language of training, guidance, growth, and research experience. The words sound noble, but when you examine the structure, you see something different.

The professor is both a guide and an evaluator. They approve graduation, write recommendation letters, influence paper authorship order, secure research funding, allocate tasks, and control lab resources. Meanwhile, graduate students are both learners and research workers, but lack the power to negotiate the conditions of their labor.

You could say that at a company, if you don't like it, you can quit. That's not easy either, but at least when you leave, the relationship ends. Graduate school is more subtle. If you leave, years of time, papers, degrees, and career paths are shaken along with you.

That's why the most powerful control mechanism in graduate school isn't salary.

It's the degree.

The moment a degree becomes a hostage, a student is no longer a researcher but a highly skilled worker awaiting graduation approval.

Not dragged, but sold a dream

Graduate students usually aren't forced into this situation. Rather, they're sold a dream.

You have talent for research. Let's write a good paper together. If you do a PhD, doors will open. You'll learn a lot in our lab. I think you'll do well.

Before admission, they talk about potential, and after admission, tasks pile up. At first, they talk about research, but later there are reports, schedules, experiments, and publication metrics.

Not all labs are like this. But this structure repeats in quite a few places, and as the exit cost—the degree—grows larger, students are cornered into enduring it. Within this, the boundary between guidance and labor control blurs, the boundary between growth and exploitation blurs, and the boundary between collaborative research and credit theft blurs.

The blurring of boundaries is the key point. If it were outright exploitation, it would be easier to escape. The problem is that it's done under the name of guidance, training, and scholarship. If the name isn't accurate, it's hard to see what's actually happening inside. When exploitation puts on sophisticated language, the person experiencing it doesn't understand it's exploitation until much later.

What was taught was not just research methodology

The bitter point is this: some professors repeat on the next generation the structure they experienced in graduate school without reflection.

What you learned in your PhD program wasn't just research methodology. You also learned how to use power, how to silence, how to make people endure, how to pass credit up and responsibility down. And some replicate that exactly. Those studying scholarship don't see the structure they belong to, talk about critical thinking to students, but remain deaf to the power dynamics of their own labs.

The title "professor" borrows the name of knowledge. That's why its power is better hidden. Power exercised through money or status is at least visible, but power exercised under the name of scholarship doesn't look like power. The most invisible power is the hardest to escape.

The era of treating people like sub-agents

The sense of déjà vu I felt while managing AI agents comes from here.

I delegate work to agents. I entrust implementation to one, review to another, have one find prior materials, have one analyze failure causes. When results are off, I ask again; when there's no basis, I demand it; I cross-verify by throwing the same problem to different agents.

This structure felt so much like a research lab. A professor doesn't conduct every experiment directly. They pose questions, divide problems, delegate to students, and review results. A good professor develops students through this process; a bad professor consumes students through this structure. AI orchestration technically resembles that structure.

But there's a crucial difference.

AI agents are not people. They don't need degrees, don't wait for graduation, don't need recommendation letters, can be called at night, can be sent back to redo things, and can be turned off if unsatisfied.

So this analogy is funny and bitter at once.

Previously, people were treated like sub-agents. Now we can actually orchestrate real sub-agents.

Behind knowledge production has always been invisible labor

The history of modern science is usually written as the history of geniuses. A few names stand in front; the rest are erased. But knowledge production has never been achieved alone. There was always labor from assistants, disciples, technicians, calculators, and experimental workers.

This isn't sentimental argument—it's a research field in the history of science. Since Steven Shapin's 1989 paper "The Invisible Technician," which examined the unnamed assistants in 17th-century laboratories, history-of-science scholarship has steadily tracked this invisible labor. The hands that actually performed experiments, the people who calculated data, the technicians who managed specimens. Their labor didn't remain as names in papers. Names went up, repetitive labor stayed down; credit went up, failure and all-nighters stayed down.

Modern graduate school has maintained that ancient structure by combining it with the institutions of degrees, papers, research funding, recommendation letters, and authorship. So it might not be coincidence that research labs came to mind with AI agents. Knowledge production has always needed a structure that divides tasks, reviews results, and makes things repeat. The problem was who held power in that structure, who did the labor, who took credit, and who bore responsibility.

Ultimately, it's a problem of asymmetry. Authority under the name of guidance, labor under the name of training, endurance under the name of growth, authorship asymmetry under the name of collaborative research, silence under the name of scholarship. When these words accumulate, exploitation wears quite a noble face.

Graduate school was originally a human agent orchestration system

At first, I thought AI agents looked like graduate students. But thinking more carefully, the direction might be reversed.

It's not that AI agents look like graduate students.

Graduate students have always been called upon, deployed, evaluated, and consumed like agents within a knowledge production system.

AI orchestration didn't mimic the graduate school lab. The graduate school lab was originally a human orchestration system. A structure that divides problems, calls upon subordinate workers, evaluates results, makes them redo things, and makes final judgments from above. This structure existed in labs, in companies, and now exists in AI orchestration.

The difference is whether the subordinate workers are human or software.

This difference is not small. AI doesn't need to take growth with it, so consuming it raises no ethical problems. In that sense, this transition removes people from the place where they were being worn down. But simultaneously, the same question moves up one level.

Have we eliminated asymmetry, or just changed the target of asymmetry?

If we're giving AI the same instructions we once gave people, we haven't created a new structure for knowledge work. We've just inserted a new subordinate worker into an old structure.

Technology illuminates old structures

AI agents look like completely new technology. But as you use them, you see old structures again. Looking into the future of technology, you see the past of knowledge production.

We're not orchestrating agents for the first time now. We've been orchestrating people like agents for a long time; we just called it training, guidance, and growth.

So was the past graduate school structure a necessity of knowledge production, or a result of people being treated like tools? There's one criterion that separates them: Do subordinate workers take away not just results but also growth, or do they send results upward while being consumed themselves?

Good labs were the former, bad labs were the latter. And what made that difference wasn't technology but the attitude of those holding power.

This point is crucial. Now that we're orchestrating AI, for the first time, we're massively experiencing what it's like to stand in the position above. The position of dividing problems and throwing them, evaluating results, and demanding a redo if unsatisfied. We do things to AI that guilt prevented us from doing to people. And the attitude we learn this way stays with us.

Technology removed people from the place where they were being worn down, but it didn't change the attitude of those standing above. That's still each person's responsibility. The way you give work to AI ultimately reflects the way you give work to people. Hands accustomed to treating subordinate workers as consumables will move the same way when the subordinate becomes human again.

Good mentors still exist and deserve respect. But the existence of good mentors doesn't justify bad structures. And what created those good mentors wasn't the system but individuals' choice to stand above without consuming those below them.

Only now, orchestrating not people but agents, do we clearly see the operation of power that had been hidden for too long under the name of scholarship. Now that we see it, it's our turn to decide what attitude to take standing in that position. Whether toward people or agents.