The Class of Cognitive Surrender

Dear young comrades

On the morning of 3 November 2026 Addy Osmani published a long post on X: "Critical thinking when agents do the work". More than two million people read it in a day. Osmani worked for years on the Google Chrome team and now works on Claude Code at Anthropic. He is one of the most widely read engineers in the software world. So this warning comes not from someone criticising artificial-intelligence agents from the outside, but from inside the company that builds them.

We agree with an important part of what Osmani says, and we find the practices he proposes valuable. But the warning is set up as a problem of attention and discipline, that is, as the individual's relation to himself. We propose to take the same fact two steps further. First, philosophically. Handing thinking over to someone else is the subject of a long argument that runs from Plato to Aristotle, from Spinoza to Kant, from Hegel to Marx and Engels, from Lukács to Gramsci, and from Vygotsky to Ilyenkov. That argument ties three questions together: how knowledge is acquired, how the human being is formed as a thinking being, and in whose hands the capacity to think is gathered in society. Then, in class terms. From whom thinking is taken, to whom it is transferred, and who pays the price are questions of property and of the labour process.

As usual, the summary first, then the detail.

---

First, the Summary: The Warning and Our Reading, in Sixteen Theses

What does Osmani say?

1. As models improve, obvious errors grow rarer. The outputs are polished, fluent and mostly correct. So they do not ask for a second look, and attention slackens.
2. There is a ladder of three steps: cognitive offloading (you hand over the work and remain the owner of the result), deferral (the tool's output becomes a new input, not a hypothesis to be tested), cognitive surrender (the output quietly becomes the person's own judgement). As trust or laziness grows, people slide downwards.
3. The bottleneck has shifted from writing code to being sure that you are right. He calls the gap between what the agents produce and what a person can stand behind the orchestration tax.
4. The prescription is individual: form a hypothesis before you ask, ask what a green test does not prove, have the agent list its decisions, write code without an agent once a week, and slow down when you need to. You can hand over the work. You cannot hand over accountability.

Our philosophical reading:

5. Thamus's criticism of writing is the earliest known criticism of cognitive offloading. Knowledge that has been externalised gives only the appearance of wisdom. But the history of writing showed that the problem does not lie in the tool itself. It lies in whose monopoly the tool is.
6. Aristotle divides knowledge into kinds. Beside knowledge that can be passed on as a rule, there is practical wisdom (phronesis), gained only through experience, which lets a person choose what is right in a concrete case. What Osmani means by "you cannot hand over accountability" is precisely this power of judgement, which cannot be handed over.
7. Osmani's diagnosis of "trust or laziness" has the same structure as Kant's linking of immaturity, in 1784, to laziness and cowardice. Both tie a true fact to the individual's morals. Marx's reply is this: who does the thinking in whose place is a question of the social division of labour.
8. For Spinoza, freedom is to grasp the causes that determine us. The person who lives through what he has not grasped is passive. The person who grasps it is active. Cognitive surrender, in Spinoza's language, is consenting to live by inadequate ideas.
9. In Hegel's dialectic of master and slave, it is labour that forms consciousness. The engineer who has agents do the work thinks he has taken the master's place, but because he has lost the formative labour, he is the one who loses. Osmani's observation that "thinking used to happen in the gap between question and answer" describes exactly the loss of this labour.
10. For Engels, the hand is not only the instrument of labour. It is labour's product. For Vygotsky, tools develop the mind only when they are internalised. For Polanyi, most of what we know cannot be put into words and is learned only by doing. These three thoughts build the epistemology of the oversight paradox: you cannot oversee without doing, because the knowledge that makes oversight possible is born inside the doing.
11. In the "Fragment on Machines" in the Grundrisse, Marx foresaw a stage in which knowledge passes into fixed capital and the worker is drawn to the side of the production process, as its watchman and regulator. The engineer who "oversees" the agents is this figure today.
12. Lukács's concept of reification is the deepest account of cognitive surrender: the "contemplative stance" of someone who looks at the product of his own labour from the outside, like a spectator. Surrender is not a failure of attention. It is the reflection, in consciousness, of a reified labour process.

Our class reading:

13. The advice to "slow down" is written for someone who can set his own pace. Where the pace is set by capital, slowing down is not a preference. It is a matter for bargaining. The gain of the orchestration tax is written to the company. The risk is written to the worker.
14. What Anthropic calls the "oversight paradox" is the form, in mental labour, of the process Braverman called deskilling: mental Taylorism. In Marx's language, the real subsumption of mental labour under capital.
15. Osmani's four separate warnings match, one for one, the four forms of alienation Marx described in 1844: from the product, from the process, from one's own capacity to think, and from other people.
16. "You cannot hand over accountability" is true, but responsibility without authority is the definition of the proletarian position. The dumping of blame, in automated systems, onto the nearest human being is called the "moral crumple zone". Our proposal is to turn Osmani's individual practices into collective rights. That happens through organising, not through discipline.

---

What Does the Warning Say?

The gap between question and answer

Osmani begins with an observation. Once, most of the work of thinking happened in the time between question and answer. We searched, tried, went wrong, and built again. That time is now measured in seconds. And because obvious errors grow rarer as the models improve, the outputs do not say "look again". Osmani says that in his daily work nothing is left that forces him to think critically.

One detail is striking. Osmani says that agent workflows are pleasant. On a heavy day it is a relief to receive an answer that is fast, fluent and mostly correct. Attention slackens easily for exactly this reason. This is directly related to what we wrote on why we deceive ourselves (in Turkish). Fluency takes the place of truth.

Offloading, deferral, surrender

Osmani's range:

• Cognitive offloading: You give the work to artificial intelligence, but you remain the owner of the result. Osmani says he does this all day, and that everyone should.
• Deferral: The tool's output becomes your new input. The result is no longer a hypothesis to be tested.
• Cognitive surrender: The output of artificial intelligence quietly becomes your judgement. You do not even feel the need to check.

The concept of "cognitive surrender" comes from a study by Steven Shaw and Gideon Nave of Wharton on how artificial intelligence reshapes human reasoning. The point Osmani stresses is that the slide is not noticed. Each step is small and reasonable. You read the previous answer carefully, you glanced at the next, you looked only at the skeleton of the one after that. Then, when someone asks about a line, you cannot explain it even with the agent's help.

The accumulation of this is called cognitive debt, and its counterpart on the code side comprehension debt. The codebase looks clean, the tests pass, but no one any longer knows how the system works. This appears only when the system has to be changed.

The bottleneck has moved

For Osmani, the bottleneck has shifted from writing code to being sure that you are right. Agents can write and run their own tests, but if the one who writes the test and the one who writes the feature carry the same false assumption, the test glows green and proves nothing. A green test confirms only one feature.

From there he moves to the orchestration tax: the gap between the work the agents produce and the work you can put your name to. If finished work piles up faster than you can review it, opening another session is not the solution. You either tighten the checks or you slow down. Osmani's summary is this. If you are producing more than you can stand behind, you are no longer reviewing. You are "hoping".

Intent debt and the oversight paradox

Intent debt: When you gave a vague request, a colleague used to say "stop, you have not thought this through". The agent does not stop. It fills the gaps and delivers something that looks finished. The aim, the limit and the reason are buried in the product without ever having been thought through. For Osmani, no one can pay this debt in your place.

The oversight paradox: Osmani takes this from research Anthropic did on its own staff. The coding skills needed to oversee Claude are the very skills that overuse of artificial intelligence can blunt. Other figures from the same research: Anthropic engineers use Claude in about 60 per cent of their work (28 per cent a year earlier), the reported rise in productivity is 50 per cent, questions are now asked of Claude rather than of colleagues, and opportunities for mentoring are shrinking. One engineer says he feels as if he comes to work every day to make himself unemployed.

The prescription

Osmani's proposals are concrete, and most of them are valuable: write your own hypothesis before you ask, test by asking "if this were wrong, what would have to be true?", hear the counter-argument from different models, define "verified" before the work, look at the commands the agent ran and at their outputs before you look at the summary, have the agent list the decisions it made and test yourself against them, write code without an agent once a week, once a month write down what you have stopped doing, put review into the work plan as a separate item, and do not publish what you will not put your name to.

The main idea: you can hand over the work, and a large part of the oversight. You cannot hand over accountability.

---

Where Do We Agree?

He describes the fact from the inside, and honestly. The marketing language of artificial-intelligence companies says "the human stays in the loop, everything is safe". Osmani describes how the human in the loop slowly leaves the loop, and does so without noticing. This matches the point we made in The Human in the Loop. A person can stay in the loop, but his place in the loop changes. He passes from being the one who does the work to the one who approves it, and from the one who approves it to the one who stamps it.

He sees that polish has stopped being a signal. Because everything comes out well formatted and sure of itself, polish is no longer a sign of quality. This meets what we wrote on Martin Fowler's "voice of the LLM". Fluency has become the sign of production, not of knowledge.

He keeps responsibility with the human being. His answer to the question "the model will be better anyway, so why should I keep the skill?" is clear. What you build, why it is valuable, whom it affects, and what you are ready to give up are your questions. This is an important stand against technological determinism. But whether these questions are really "yours", as we will see below, is a question of property.

---

The Philosophical Ground: The Long History of Leaving Thinking to Someone Else

The problem Osmani describes arrived with a new technology, but it is not a new problem. For at least two and a half thousand years humanity has argued about what it means to leave part of thinking to tools, to institutions and to other people. Following that history lets us see where Osmani is right, and where he stops halfway.

Plato and the remedy of writing

In Plato's dialogue Phaedrus, Socrates tells an Egyptian legend. The god Theuth invents writing and tells King Thamus that it is a remedy for memory and wisdom. Thamus objects. Writing does not strengthen memory. It increases forgetfulness. People will remember knowledge not from within themselves but from signs outside. They will gain not wisdom itself but its appearance. People will appear who have read much and learned nothing, who look wise and are not.

This is the earliest known criticism of cognitive offloading. What Thamus calls "the appearance of wisdom" is the same thing Osmani calls "a polished output that has not been understood". In both, externalised knowledge takes the place of internalised understanding.

But Thamus has a lack, and that lack is instructive for us. Writing did not make humanity stupid. On the contrary, it made it possible for knowledge to accumulate across generations, and for science and philosophy to exist. The problem was not writing itself. It was in whose hands writing was. In ancient Egypt writing was the monopoly of a scribe class. As literacy spread among the people, writing became an instrument of emancipation. The question is the same for artificial intelligence. Is the problem the tool itself, or the tool's being the monopoly of a handful of companies?

Aristotle: rule, craft and practical wisdom

In the sixth book of the Nicomachean Ethics, Aristotle divides knowledge into kinds. Episteme is scientific knowledge, which can be proved and taught. Techne is knowing how to make something, that is, craft. Phronesis is practical wisdom: the capacity, in a concrete situation where general rules are not enough, to see what is right and to decide accordingly.

Aristotle has an important observation. The young can become skilled in geometry and mathematics, but they cannot possess practical wisdom. Practical wisdom concerns particular situations, and recognising those situations is gained only through experience, over time. A rule can be taught. The capacity to see how the rule is to be applied in a given case cannot be taught. It is acquired by living it.

This distinction is very fruitful for understanding today. Artificial-intelligence models have gained an extraordinary power in the fields of episteme and techne. They know the rules, they apply the patterns, they produce products. What Osmani calls "not transferable" corresponds to phronesis. What should be built, why, for whom, and what one should accept giving up? Osmani's question, "would you bet on this in front of the person who will have to maintain it?", is the question of phronesis.

But Aristotle's warning gains importance a second time here. If practical wisdom is gained through experience, what happens when the experience itself disappears? Phronesis matures inside techne, while one wrestles with particular jobs, by going wrong and correcting. The person who loses his craft loses, in time, his power of judgement too. The first philosophical statement of the oversight paradox is in Aristotle.

Another of Aristotle's observations is striking for its class content. In the Politics he says that if the day comes when the shuttle weaves by itself and the lyre plays by itself, the overseer will have no need of the worker, and the master no need of the slave. For Aristotle this is a thought experiment. For us it is today's reality. The real question is this. When the shuttle begins to weave by itself, will the gain go to the weaver, or to the owner of the shuttle?

Kant: the state of immaturity, and the guardians

In 1784 Immanuel Kant, answering the question "What is Enlightenment?", begins like this. Enlightenment is the human being's emergence from his self-incurred immaturity. Immaturity is the inability to use one's own understanding without the guidance of another. Kant's slogan is Sapere aude: have the courage to use your own understanding.

There is a passage in this text that is very close to today. If I have a book that understands in my place, a spiritual adviser who has a conscience in my place, a doctor who decides my diet in my place, I need not trouble myself. So long as I can pay, I need not think. Others will take on this tiresome business in my place. Kant speaks of the "guardians" who keep this state going: authorities who make it their business to hold people in the comfort of immaturity, and who say that walking on one's own is dangerous.

It is not hard to add an item to this list today. If an agent writes code in my place, tests its own code, criticises its own work and presents me with a summary, I need not think.

There is an important likeness here. Kant places the cause of immaturity in laziness and cowardice. Osmani too says that people slide into surrender as trust or laziness grows. Both diagnoses tie a true fact to the individual's moral weakness. The solution too is individual: show courage, build discipline.

The Marxist tradition respects this text of Kant's, but adds a question. Where does the power of the guardians come from? Why are people lazy and cowardly? In Kant's time a peasant's being unable to read and write was not a moral fault. It was the result of the feudal social structure. Today too, an engineer's not finding time to read the agents' output is not laziness. It is the result of the way the labour process is organised. The guardians are no longer priests and physicians. They are the owners of the tool and the setters of the pace.

Spinoza: to be active, to be passive

A century before Kant, Baruch Spinoza had defined freedom from an entirely different place, in the Ethics. For Spinoza, freedom is not a will without causes. Everything is determined by causes, and so is the human being. Freedom is to grasp the causes that determine us.

Spinoza speaks of two kinds of idea. Adequate ideas are ideas in which we grasp a thing together with its causes. Inadequate ideas are fragmentary ideas, cut off from their causes, which reflect only how a thing appears to us, its effect, its surface. The person who moves by adequate ideas is active: the cause of what he does is himself. The person who lives by inadequate ideas is passive: he is dragged by someone else's effect.

Cognitive surrender, in Spinoza's language, is precisely passivity. The tests glow green, the summary is favourable, the output is fluent. The idea in the engineer's hands is an idea only of the appearance of the work. He does not know why it was written this way, which assumption it rests on, or where it might break. Osmani's criterion, "if you cannot explain it even with the agent's help, you have handed over ownership", is the engineering counterpart of Spinoza's criterion of the adequate idea.

What is valuable for us in Spinoza is this. To be active is not a mood or a display of will. It is a matter of knowledge. And knowledge, in Spinoza, grows stronger not alone but together with other people. For people who follow reason, he says, the most useful thing is other people who reason. In I Read It, So What? we discussed reading as the means of being active in life. In the age of agents too, the way to be active runs through grasping the causes, and through doing that together.

Hegel: labour forms consciousness

The dialectic of master and slave in Hegel's Phenomenology of Spirit explains what Osmani senses but does not name. The master has the slave work for him, and he only enjoys what is produced. The slave deals directly with the object and forms it. The conclusion Hegel reaches is surprising. In time it is the slave whose consciousness develops, not the master. Because labour, while forming things, also forms the consciousness of the one who works. Meeting the resistance of the object, going wrong and trying again, transforms the human being. The master consumes, and freezes.

Osmani's observation that "most of the work of thinking used to happen in the gap between question and answer" is exactly a description of this formative labour. The hours spent debugging, the mistakes made while reading documents, rebuilding a concept in our own words: these were not "inefficiency". They were the process in which consciousness forms itself.

In the agent workflow the engineer thinks he is in the master's place. He runs four sessions, and the agents work for him. But Hegel's dialectic works in reverse. The one who deals with the object, who forms it, who goes wrong, is no longer the engineer. The engineer falls into the position of the master who benefits from the results but has lost the formative labour. The oversight paradox is the name of this. You cannot oversee without forming.

There is a very important difference. Hegel's master really was a master. Today's engineer is not a master. He does not own the tool, he does not set the pace, the gain does not go to him. So he both loses the formative labour and does not possess the master's privileges. He combines the worst sides of both positions.

Engels: the hand is both the instrument and the product of labour

In the essay he wrote in 1876, "The Part Played by Labour in the Transition from Ape to Man", Friedrich Engels puts forward a materialist thesis. The hand is not only the organ of labour. It is also labour's product. The human hand developed, across thousands of generations of work, by adapting to ever finer tasks. The hand that made Raphael's paintings, Thorwaldsen's sculptures and Paganini's music is the product of the long history of labour. For Engels, labour, language and consciousness developed together.

This thesis sets on a material ground what Hegel sensed in an idealist form. Skill, capacity, even the capacity to think, are not fixed properties the human being possesses at birth. They are capacities formed inside activity, and lost without activity. The unused hand thins. The unused capacity blunts.

What Anthropic calls the "oversight paradox" is, seen from this angle, not a surprise. It is a materialist necessity. The skill of writing code developed by writing code, the intuition of debugging by debugging. When this activity disappears, the skill that was its product does not remain. Osmani's warning that "the skills you stop exercising disappear faster than you think" is the expression, at the individual level, of Engels's thesis.

Marx: the general intellect, and the watchman of the production process

In the section of the Grundrisse, written in 1857-58, known as the "Fragment on Machines", Marx makes an extraordinary prediction. As large-scale industry develops, social knowledge, that is science, technique and accumulated skill, is increasingly objectified in machines, in fixed capital. Marx calls this the general intellect. At this stage the worker ceases to be the real subject of the production process, moves to its side, and becomes its watchman and regulator.

Read this beside Osmani's post. The engineer who runs the agents, watches what they produce, checks the evidence and approves it is exactly the watchman and regulator Marx described. Osmani says this openly too. The human mental model is no longer in the middle of the process. It is at the beginning and at the end.

Marx's emphasis is this. The transformation is not bad in itself. The accumulation of social knowledge in machines carries the potential to free humanity from necessary labour. But inside capitalist relations this accumulated knowledge confronts the worker as an alien power, the power of capital. The general intellect is everyone's product, but its property is capital's.

This is what is happening in artificial-intelligence models today. The open code written over years by millions of developers, the questions answered, the documents written, have been transferred into the models. This is the process we called the expropriation of the general intellect. Now the same developers work as the watchmen of a tool made from their common knowledge, and this tool blunts their skill.

Lukács: reification and the contemplative stance

In History and Class Consciousness, in 1923, György Lukács extends Marx's concept of commodity fetishism and develops the concept of reification. In capitalist production the labour process is broken into parts, made calculable, and the worker becomes a part of this process. The counterpart in consciousness is what Lukács calls the contemplative stance. The worker looks at his own activity from the outside, as a spectator, as if it were something operating according to a law independent of himself. He ceases to be active, and watches what happens.

This is the deepest account of cognitive surrender. The engineer Osmani describes watches what the agents produce, like a spectator. The tests glow green, the summary is favourable, the skeleton looks reasonable. He looks at his own work from the outside and accepts it like a natural event. In Lukács's eyes this is not a failure of attention. It is the reflection, in consciousness, of a reified labour process. So long as the process places the individual in the spectator's position, however disciplined the individual is, he has to wage a constant struggle against the pull of that position.

Osmani's warning that "by default nothing will do this for you, you have to place critical thinking in your day" diagnoses this pull correctly. But he still looks for the solution in the individual's will. Lukács's answer is different. Reification is overcome only by a collective practice and a class consciousness in which the worker grasps himself as the subject of the production process.

Gramsci: every person is a philosopher

In the Prison Notebooks Antonio Gramsci says that every person is a philosopher. Everyone carries a view of the world inside language, common sense and daily experience. But this "common sense" (senso comune) is mostly fragmentary, contradictory, and filtered from the ideas of the ruling classes. The work of philosophy is to turn this common sense into a critical consciousness.

We have to see that the age of agents produces a new common sense. Model outputs offer a "reasonable view" that is filtered from the average of the internet, shaped by the training choices of monopolies, fluent and sure of itself. Cognitive surrender is the adoption of this new common sense without questioning it. In Gramsci's eyes this is a problem of hegemony. People's way of looking at the world is fed not by their own experience but by a source someone else produced and someone else controls.

Gramsci's solution is the "organic intellectuals": people who come from inside their own class and turn experience into critical consciousness. Today the counterpart of this in computing labour is the workers who question the agents while using them, share that questioning with their friends, and turn it into a ground for organising.

Vygotsky and Ilyenkov: thinking is an activity

The Soviet psychologist Lev Vygotsky showed that the human mind develops through tools. But he makes an important distinction. Tools and signs, that is language, writing and number systems, are first used outside, among people, inside a social relation. Then they are internalised and become the mind's own way of working. The child first thinks by speaking aloud, then this speech turns inward and becomes inner speech, thought. Every higher mental function, for Vygotsky, appears first on the social plane, then on the individual plane.

Vygotsky's second important concept is the zone of proximal development: the area between what a child can do alone and what he can do with the help of an adult or a more experienced friend. Learning happens in this area. But the aim of the help is to withdraw over time, and to make it possible for the child to do the same work alone. Help that never withdraws produces dependence, not learning.

This concept shows the junior engineer's situation in the age of agents very clearly. The agent can be a helper in the zone of proximal development. It helps the junior do what he could not do alone. But if the agent never withdraws, that is, if the junior never has to do the same work without an agent, internalisation does not take place. The tool stays outside. It does not become part of the mind. Osmani's question, "how lost would you be if there were no artificial intelligence?", measures exactly this. When a tool that has not been internalised disappears, nothing is left behind.

A finding from Anthropic's research gains importance again here. Questions are now asked of Claude, not of colleagues. For Vygotsky the first plane of learning is social: thinking together with a more experienced person. When this plane is withdrawn, learning itself changes.

The Soviet philosopher Evald Ilyenkov takes this argument one step further. For Ilyenkov, the "ideal", that is thought, meaning and concept, is not inside the individual's head. It exists in the objectified forms of social activity, in tools, in language, in institutions. The individual learns to think by reproducing these objectified forms through his own activity. Learning to use a hammer as a hammer is to bring to life, in one's own activity, the knowledge humanity has accumulated in the hammer.

Artificial-intelligence models, in Ilyenkov's language, are an extraordinarily dense objectification of humanity's activity of thought. The code, the writing and the argument of millions of people have been gathered in one tool. But something unseen in history is happening here. This objectified thought works without requiring the person who uses it to bring it to life again in his own activity. To use a hammer, one has to learn to use a hammer. To use an agent, one does not have to learn to write code, at least at the start. Social knowledge is sold as a service, without returning to the individual's activity. The chain of internalisation breaks for the first time.

Polanyi and Ryle: knowledge that cannot be said

The Hungarian-born philosopher of science Michael Polanyi has a famous sentence. We know more than we can say. Polanyi calls this tacit knowledge. We can recognise a face but cannot tell how we recognise it. We can ride a bicycle but cannot turn into a rule how we keep our balance. An experienced physician senses, looking at a patient, that something is wrong, but cannot fully explain this intuition.

Gilbert Ryle makes a similar distinction. Knowing that something is so (knowing that a proposition is true) and knowing how to do something are not the same thing. Knowing the rules of chess is not knowing how to play chess well.

A senior engineer's saying "there is a smell in this code" is exactly tacit knowledge. It is the intuition, which cannot be turned into a rule but is real, of someone who has met hundreds of errors and tried to keep hundreds of systems standing. In Osmani's example of a 600-line pull request, it is this intuition that would catch the fine change at the transaction boundary, or the default value that has been reversed.

Polanyi's most important finding is this. Tacit knowledge cannot be passed on by being written down. It is passed on only by doing together, inside the relation of master and apprentice. So the oversight paradox is not only a loss of skill. It is a break in transmission. The knowledge that makes oversight possible is tacit. Tacit knowledge is learned only by doing. When the doing passes to the agent, tacit knowledge is neither produced nor passed on. This is not a loss that can be made good by an individual exercise. Unless the social relation in which master and apprentice work together is preserved, the passage of knowledge from generation to generation is cut.

The practical criterion: the green test and social practice

The second of Marx's Theses on Feuerbach says this. Whether human thought attains objective truth is not a theoretical question but a practical one. The truth of thought is proved in practice.

Osmani's emphasis on verification is close to this principle. Giving the agent the means to verify its own work, testing the result, looking at the evidence: these are the testing of thought in practice. But Osmani himself also sees how this practice narrows. The question "is there evidence?" turns into the question "did the test glow green?"

Marx's practice, though, is not a narrow technical check. It is a social activity. Whether code is "correct" is tested not only by the test's passing, but in the user's life, in the work of the worker who will maintain it, and in society's need. Osmani's criterion, "would I bet on this in front of the person who will maintain it?", is a step in this direction. We take one step further. Can I defend this code in front of the people who will use it, be affected by it, and live its results?

The third thesis connects here too. The educator himself has to be educated. By whom, and under what conditions, is the engineer who oversees the agents trained? If the ladder that trains the overseer is pulled up, oversight too is emptied in the end.

Synthesis: the three layers of thinking

When we set these thinkers side by side, what Osmani sees as a single problem turns out to have three layers.

The epistemological layer: how is knowledge acquired? Aristotle, Polanyi and Ryle show that the most valuable part of knowledge, the power of judgement and tacit knowledge, cannot be passed on as a rule. It is gained only by doing. Oversight rests on this knowledge. At this layer, cognitive surrender is the disappearance of the activity in which knowledge is produced.

The ontological layer: how is the human being formed as a thinking being? Hegel, Engels, Vygotsky and Ilyenkov show that the capacity to think is not given. It is formed inside activity. The hand is the product of labour. The mind is the product of internalised tools. At this layer, cognitive surrender is the cutting of the human being off from the activity that forms him.

The social layer: in whose hands is the capacity to think gathered? Thamus in Plato, Kant's guardians, Marx's general intellect, Lukács's reification and Gramsci's hegemony show that where knowledge accumulates, and under whose control, is a question of power and property. At this layer, cognitive surrender is the turning of social knowledge into the capital of a handful of companies, and the turning of people into the watchmen of this capital.

The first two layers explain why the problem is real and deep. The third layer explains the problem's direction. Why are these tools organised in a way that makes people dependent, rather than in a way that develops their capacity? Why does subscription come forward instead of internalisation, and watchmanship instead of craft? This comes not from the nature of the technology, but from the property relations inside which the technology is set.

Thinker	Basic concept	What it says to the age of agents
Plato (Thamus)	The appearance of wisdom	Polished output can take the place of understanding. But the question is the monopoly of the tool.
Aristotle	Phronesis, practical wisdom	The power of judgement cannot be handed over, but if experience disappears, judgement blunts too.
Spinoza	The adequate idea, being active	Living with an output whose cause you have not grasped is passivity.
Kant	Immaturity, guardians	The comfort of leaving thinking to someone else produces new guardians.
Hegel	Master and slave, formative labour	It is the one who does the work, not the one who has it done, who gains consciousness.
Engels	The hand is the product of labour	An unused capacity blunts. The oversight paradox is a materialist necessity.
Marx	General intellect, watchman and regulator	Social knowledge passes to capital, and the human being becomes the watchman of the process.
Lukács	Reification, the contemplative stance	Surrender is becoming the spectator of one's own work.
Gramsci	Common sense, hegemony, the organic intellectual	Model output produces a new common sense. Critical consciousness develops through organised practice.
Vygotsky	Internalisation, proximal development	Help that never withdraws produces dependence, not learning.
Ilyenkov	The objectification of the ideal	Social thought is sold without returning to the individual's activity.
Polanyi and Ryle	Tacit knowledge, knowing how	The knowledge that makes oversight possible is passed on only in the relation of master and apprentice.

---

The Class Reading: From Whom Is Thinking Taken, and to Whom Is It Given?

Formal and real subsumption

In the unpublished "Chapter Six" of Capital, Marx develops two concepts. In formal subsumption, capital takes over the existing labour process as it is. The craftsman works as before, but now for a boss. In real subsumption, capital reorganises the labour process itself. The work is divided, mechanised, and the pace and the method are set by capital. The worker is no longer the owner of the process. He is a part of it.

The labour of software stayed for a long time at the stage of formal subsumption. People worked for a boss, but how the work was done was left largely to the engineer's knowledge and preference. The programmer's relative autonomy, high wages, and identity as "a professional, not a worker" came from here. Agent workflows are the instrument of the passage to real subsumption. The content of the work is tied to the model, the method to the agent framework, the pace to the volume of output. What remains in the engineer's hands is the role Marx called watchman and regulator.

The technology worker's position between the mental proletariat and the labour aristocracy (in Turkish) is changing fast in this passage. Osmani's warning can be read as the first shock, in consciousness, of this passage.

To whom is "slow down" said?

Osmani proposes slowing down if finished work piles up faster than it can be reviewed. A correct proposal. But who sets the pace?

Let us think of an engineer working in a software company. His manager has read that productivity rose by 50 per cent with agents. Quarterly targets have been raised accordingly. The team has been shrunk, because the same work can now be done with fewer people. Microsoft's having artificial intelligence write forty per cent of the code while cutting a quarter of the staff was the most concrete example of this. Does this engineer have the possibility of saying "I will slow down, because I am producing more work than I can stand behind"?

Marx describes this as the intensification of labour. When the working day cannot be lengthened, the labour spent in the same time is increased. The worker tied to the machine works at the machine's speed. The engineer who manages four agent sessions in parallel is the counterpart, in mental labour, of the same logic. As Osmani says, however many agents you run, there is one pair of eyes looking at the outputs. The authority to set the speed of those eyes is not with the owner of the eyes.

The class content of the orchestration tax comes into the open here. The gain in speed is written into the company's balance sheet. The risk of work that cannot be reviewed stays on the worker's shoulders.

Responsibility without authority, and the moral crumple zone

The strongest sentence in Osmani's post is that accountability cannot be handed over. We take this seriously, and we add a question. Can responsibility be loaded onto someone who has no authority?

The technology anthropologist Madeleine Clare Elish gave the name "moral crumple zone" to a pattern in automated systems. The crumple zone in cars protects the driver by absorbing the blow of the collision. In automated systems too, when something goes wrong, the blame is dumped not on the one who designed or operated the system, but on the human being standing nearest the process. In the autopilot's error the pilot is blamed, in the crash of the driverless car the safety driver. The human being becomes the buffer that absorbs the system's moral and legal blow.

The engineer in agent workflows is a new moral crumple zone. He did not set the pace, he did not choose the tool, he did not decide to shrink the team. But when the transaction-limit error in a 600-line pull request reaches production, the question is clear. "Who approved this?"

Responsibility without authority is the definition of the proletarian position. The worker is held responsible for the results of a process whose decision he did not make. Osmani's principle "you cannot hand over accountability" is emancipatory when it comes together with the authority to oversee. When it is left without authority, it is an instrument of discipline.

Four alienations, four warnings

In the Economic and Philosophic Manuscripts of 1844, Marx describes four forms of the alienation of capitalist labour. Osmani's warnings overlap with these in a striking way.

Marx's form of alienation (1844)	Its counterpart in Osmani's warning (2026)
Alienation from the product of labour: What the worker produces confronts him as an alien object.	Comprehension debt: The codebase is clean, the tests pass, but no one understands it. The engineer cannot explain the code he himself approved.
Alienation from the labour process: Work is not the worker's own activity. It is an activity imposed from outside.	The orchestration tax: The pace is set by the volume produced. The human being is taken out of the middle of the process and placed at its beginning and its end.
Alienation from species-being: The conscious, creative activity that makes the human being human is reduced to a means of subsistence.	The oversight paradox and cognitive surrender: The capacity to think itself blunts. The question "what have I stopped doing?" is the question of this loss.
Alienation from other people: People relate to one another as rivals and as instruments.	The shrinking of mentoring: According to Anthropic's research, questions are now asked of Claude, not of colleagues. The learning bond between junior and senior weakens.

This table shows one thing clearly. The things Osmani sees as four separate technical problems are four faces of the same social relation. So they cannot be solved by separate individual prescriptions.

Mental Taylorism

The essence of Frederick Taylor's method, as Harry Braverman tells it in Labor and Monopoly Capital, is to take knowledge out of the worker's head and gather it in the hands of management, and to separate conception from execution. The master is the person who knows the whole of the work. Taylorism transfers the master's knowledge to cards, to instructions and to machines. In the master's place there remains a worker who repeats a single movement.

What is happening in software is the mental version of this. But there is an interesting reversal. In classical Taylorism, conception stayed with management and execution with the worker. In the agent workflow, execution passes to the model, and Osmani tells the engineer "the intent is yours, the conception is yours". At first glance this looks like the worker's winning conception back.

But it is not. Because the experience of execution that feeds conception is disappearing. As Hegel showed, conception develops by wrestling with the object. A conception cut off from execution is gradually emptied. What Osmani calls "intent debt" is the result of this emptying. A mind that does not meet the resistance of the object in execution cannot clarify its own intent either. In the end both conception and execution pass under the control of the tool's owner. Mental Taylorism is the process that first transfers the master's knowledge to the machine, and then leaves the master as the machine's watchman.

Who makes the warning?

We do not doubt Osmani's honesty. But we have to see from where the piece is written. The engineer of the company that sells the agent tool describes the risk the tool creates, and shows the solution again inside the tool: self-criticism with a sub-agent, having the agent sit an examination, building the checks with the agent. These solutions can work. But when the source of the problem and the source of the solution are the same company's product, the real question is not asked. Who wants this speed, and to whom does the gain go?

This connects with what we wrote on Anthropic's public offering. The company's valuation rests on how fast and how cheaply the agents do human work. The warning to "slow down" is advice in tension with this business model. That Osmani says it is valuable. But putting this advice into effect in the workplace cannot be left to the good will of companies.

The limit of individual discipline

Osmani's prescription assumes an individual who constantly watches himself, who protects his own skill with exercise, and who sets rules against his own weakness. This resembles the Stoics' ideal of self-mastery. A valuable ideal. But under today's working conditions this ideal can also turn into a neoliberal form of subjectivity that dumps responsibility onto the individual. Protect your employability, refresh your skill in your own time, manage your own burnout.

Here we are not criticising Osmani. We are asking under what conditions the individual he describes lives. Writing code without an agent once a week, once a month writing down what you have stopped doing, having an examination prepared for every pull request: all of these take time. That time comes either out of working hours or out of the worker's private life. The way it comes out of working hours is bargaining. Its coming out of private life is a new form of exploitation.

---

Two Readings Side by Side

Osmani's individual reading	Our philosophical and class reading
Cognitive surrender is a slackening of attention formed by trust and laziness.	Cognitive surrender is the reflection, in consciousness, of a reified labour process: Lukács's contemplative stance.
Thinking used to happen in the gap between question and answer.	That gap was the labour that forms consciousness. Its loss is the freezing of Hegel's master.
The orchestration tax is the gap between the work produced and the work one can stand behind.	The gain of this tax is written to capital, the risk to labour. It is the mental form of the intensification of labour.
The solution: tighten the checks, or slow down.	The right to slow down is not an individual preference. It is a matter for collective bargaining.
The oversight paradox is the blunting of skill through overuse.	The oversight paradox is the real subsumption of mental labour under capital: mental Taylorism.
Write code without an agent once a week.	Time for learning and for protecting skill must have a place inside paid working time.
You cannot hand over accountability.	Responsibility without authority is the moral crumple zone. Responsibility must come together with the authority to oversee.
No one can pay your intent debt in your place.	If management, which sets the aim of the work, sets the intent, the debt must be shared too.
What you build, and why it is valuable, are your questions.	These questions can be the questions only of those who have a say over the means of production. This is a question of property.

---

So What Is to Be Done?

There is no harm in adopting the practices Osmani proposes. On the contrary, they are a good beginning. The problem is that they are left to the individual's good will and spare time. Our proposal is to turn these practices into collective rules and rights.

At the level of the team:

• A limit on review capacity. The amount of work a team can review at the same time (a WIP limit) must be set openly. If agent output passes this limit, the limit is not raised. The output is slowed. The decision must belong to the team, not to engineers one by one.
• The right to say "I cannot stand behind this". A worker's refusing to approve work he does not understand, or cannot review, must not be counted as a performance problem. It must be recognised as a guarantee of quality and safety.
• A shared definition of "verified". What evidence is required for a piece of work to count as finished must be settled in advance, and in writing.
• Collective review. Critical changes must not be left to a single person's approval. Critical thinking is a practice that develops by arguing together.

At the level of the workplace:

• Time for learning and for protecting skill. The proposal to "write code without an agent once a week" must be written into working hours, not into the worker's evening. This time must have a place in the contract.
• Junior employment and mentoring. Senior engineers' mentoring time must be counted in the calculation of workload. Hiring juniors must be defended as an investment in the institution's future capacity to oversee.
• Changing the measures of productivity. Productivity must be measured not only by the code produced or the number of jobs closed, but by the error rate, the burden of maintenance, the people trained, and reliability.
• Matching responsibility with authority. In an error, the worker's responsibility must be assessed together with the responsibility of those who made the decisions on pace, on the choice of tool, and on the size of the team.

At the level of society:

• Keep the general intellect a commons. The knowledge artificial intelligence feeds on is society's common product. Open documentation, open teaching material, open code and free models (in Turkish) let this knowledge accumulate in society's common memory, not in the models of companies. The answer to Thamus's question is not to ban the tool. It is to socialise the tool's property.
• A culture of reading and thinking. Osmani's proposal of active reading is part of the relation between reading and being active that we discussed in I Read It, So What?. Kant's sapere aude gains meaning today only in the practice of people who think together.

---

Concrete Tasks for Computing Workers

1. Make the orchestration tax visible. For one week, note what the agents produced and what you really reviewed. Talk about the gap in the team meeting.
2. Make the right to say "I cannot stand behind this" a team rule. Propose a written team agreement.
3. Set up regular pair-working hours with your junior friends. Without an agent, together, by asking.
4. Question the measures. When your manager says "productivity has risen", ask about the error rate, the changes that were reverted, and the burden of maintenance.
5. Set up reading groups. Read together the texts we named in this piece, from Kant's short essay to Engels's essay on labour, from Vygotsky to Braverman, and argue them against your own experience of work. Theory is the means of turning experience into consciousness.
6. Share the experiences. Discuss how cognitive surrender is lived in your workplace not as an individual weakness but as a common problem. The small nuclei we proposed in Organising 2.0 are a suitable ground for this.
7. Carry it onto the union agenda. Review time, learning time, junior employment, how artificial-intelligence tools are brought into the workplace, and how responsibility is shared when there are errors, can be clauses of a collective agreement. For that, one has to organise in a computing union.

---

Dear comrades

Addy Osmani ends his post with a few simple pieces of advice. Think independently, ask better questions, take responsibility for the result. We agree with this advice. But we add something. Independent thinking has material conditions. For someone who cannot spare time to think, who cannot set his pace, who is afraid of losing his job when he says "I am slowing down", critical thinking becomes a luxury.

Thamus feared that writing would give only the appearance of wisdom. Aristotle said that the power of judgement matures only through experience, Spinoza that freedom is to grasp the causes. Kant warned against the guardians who hold people in the comfort of immaturity. Hegel showed that labour forms consciousness, Engels that the hand is the product of labour, Marx that social knowledge turns into the power of capital, Lukács that the worker turns into the spectator of his own life. Vygotsky showed that tools develop the mind only when they are internalised, Polanyi that the most valuable knowledge is passed on only by doing together. Gramsci said that every person is a philosopher, and that this philosophy can turn into a critical consciousness only through an organised practice.

All of these thinkers have a common lesson. Thinking is not a product. It is an activity. The person from whom the activity is taken cannot be the owner of the product either.

The machine is new, and the chain is old. Taylor, who transferred the master's knowledge onto cards, and the companies that transfer the general intellect into models, ask the same question. In whose hands is the knowledge, under whose control is the work, for whose benefit is the gain? The strongest antidote to cognitive surrender is not the will of persons one by one. It is organised workers who together demand time to think, the right to learn, and the authority to oversee that comes together with responsibility.

We cannot hand over responsibility. That is true. But we should not take on responsibility either without taking its authority. We should complete the courage to use our own understanding with the organising of using our understanding together.

---