From human computers to Michelin kitchens: why every era of computing has really been a redesign of the organization, why Aristotle and Kant drew our agent architectures long before we built them, and why the one question that never gets automated is how we know the work is true.
Philadelphia, 1943. A young woman sits at a desk calculator in a room at the Moore School of Electrical Engineering. In front of her is one line of one firing table for one artillery piece: where a shell will land, given this charge, this elevation, this air. To produce that single trajectory, she will work for roughly forty hours, stepping through a differential equation by hand, each result feeding the next, each digit copied onto the next row. She is not called a mathematician. She is called a computer. Somewhere inside those forty hours is a part of the work nobody photographs. She stops. She takes the differences between successive rows and looks for the one that breaks the pattern. She checks. The shell does not care how elegant the method was. It cares whether the number was right.
Tonight, somewhere else, a single engineer runs a hundred agents. They refactor, write tests, open pull requests, close tickets, open new ones. A slide that has been passing around this year draws the arc between those two rooms in four panels — human computers, programming languages, coding agents, software factories — and captions the last one plainly: engineers decide what to build, agents decide how.
The slide is right. But it tells the story as a story about machines, and it is not one. It is a story about organizations: how we cut thinking into pieces, hand them out, and decide whom to believe when the pieces come back. The machine has changed four times. That question has not changed once. It is about two and a half thousand years older than computing.

The first software factory was made of hairdressers
Look again at the first panel of that slide: a few figures at the top, more beneath them, rows of calculators at the base. It is labeled the 1940s. The shape is older. In the early 1790s, revolutionary France wanted to make the metric system real, which meant new logarithmic and trigonometric tables at a precision nobody had attempted — logarithms of every number up to two hundred thousand, carried to fourteen places and beyond. The engineer charged with the job, Gaspard de Prony, had been reading Adam Smith.
Smith opened The Wealth of Nations in 1776 with a pin factory. One man working alone might make twenty pins a day; ten men who split the trade into its eighteen operations make forty-eight thousand. De Prony drew the reckless and correct conclusion. He would manufacture logarithms the way one manufactures pins.
He built three tiers. At the top, a handful of eminent mathematicians — Legendre among them — chose the formulas. In the middle, skilled calculators turned those formulas into numerical procedure: starting values, schedules of differences, the forms to be filled in. At the bottom sat around ninety workers who needed to know nothing but how to add and subtract, and who followed the procedure exactly. Many of them were hairdressers, men whose trade had collapsed along with the aristocracy and its wigs, and whose habits of patient precision turned out to be precisely what the bottom tier required.
Planner. Decomposer. Executor. The top decides what is to be computed, the middle decides how, the bottom does it. This is not an analogy for an orchestrator with subagents. It is one, instantiated in people, more than two centuries before anyone typed the word agent. Charles Babbage understood what he was looking at. Around 1821, checking two independently computed sets of tables with John Herschel and finding error after error, he is said to have exclaimed that he wished to God these calculations had been executed by steam. In On the Economy of Machinery and Manufactures (1832), he made de Prony’s workshop his central proof that the division of labor applies to mental operations as surely as to mechanical ones — and then drew the inference that changed everything. The work of the bottom tier had already been made so mechanical that a machine could do it. The Difference Engine was designed to replace the hairdressers.
Here is the first law of this entire history, and it holds at every panel of the slide. Mechanization never begins at the top. It begins wherever the work has already been reduced to procedure. The organization performs the decomposition first; the machine then absorbs the tier the organization has already made mechanical. Automation is the shadow of a prior act of management.
The women who were not allowed to see the machine
Aberdeen Proving Ground needed firing tables faster than any room of human computers could produce them. ENIAC was built to absorb that tier. Then something happened that the planners had not quite anticipated. The machine did not remove the computers. It moved them up.
Six women from the computing corps — Kay McNulty, Jean Jennings, Betty Snyder, Marlyn Wescoff, Ruth Lichterman and Fran Bilas — were chosen to make ENIAC do the work they had been doing by hand. No programming languages, no manuals, no instruction codes. They worked from wiring and logic diagrams, and for a time did so without the security clearance to enter the room where the machine stood. Kathy Kleiman spent decades recovering their story for Proving Ground, the book cited in the corner of that very slide. They invented a craft: translating a problem into a machine’s physical configuration.
Notice the motion. The human computers did not vanish. They stopped executing the how and started specifying it. That is exactly the promotion the slide now offers engineers, and it has happened before, to specific people with names. Notice the other half too. When ENIAC was unveiled to the public in February 1946, the women were not introduced. They were trusted with the work, not the credit. Institutions have always been clearer about what they will delegate than about whom they will recognize. Hold on to that; it comes back.
The second panel — compilers and high-level languages — continues the same motion one tier higher. Grace Hopper recalled that when she had one of the first working compilers in the early 1950s, nobody would touch it; people were certain a computer could do arithmetic but could not write programs. The objection was never really technical. It was a question of trust: will the translation from intent into machine be faithful? Over seventy years, the answer slowly became yes, because compilers became deterministic, specified, and testable. We stopped reading assembly because we could demonstrate that we no longer had to. A coding agent is a compiler that is not deterministic. That single fact is why the last two panels are so much harder than the first two.

Theoria, praxis, poiesis
Aristotle divided all thinking into three kinds. Theoretical thought concerns what cannot be otherwise. Practical thought concerns praxis — action whose end is the doing itself, acting well. Productive thought concerns poiesis — making, whose end lies outside the activity: the house, the ship, the cure. Each had its excellence. Episteme, knowledge demonstrable from first principles. Techne, craft, a stable capacity to produce guided by true reasoning. Phronesis, practical wisdom, the capacity to deliberate well about what to do in this situation, with these people, now.
Anyone who builds agents recognizes the anatomy at once. The model’s latent knowledge aspires to episteme. The loop of reasoning and then acting — an architecture literally named ReAct in 2022 — is an attempt at praxis. The tool call is poiesis: the point where thought leaves the mind and alters the world, a file written, a message sent, a row changed. The frameworks we argue about are, whether they know it or not, arguments about how these three are joined.
Aristotle also asked what a tool is. In the first book of the Politics he separates instruments of production, like the shuttle, from which something further results, from instruments of action, like a garment, which are simply used. And then he writes a passage that should hang above every agent lab: “For if every instrument could accomplish its own work, obeying or anticipating the will of others, like the statues of Daedalus, or the tripods of Hephaestus … if, in like manner, the shuttle would weave and the plectrum touch the lyre without a hand to guide them, chief workmen would not want servants, nor masters slaves.” — Aristotle, Politics I (tr. Jowett)
Twenty-three centuries ago, the self-operating tool was the thought experiment that would dissolve the hierarchy of labor. Aristotle raised it in order to defend that hierarchy: the shuttle did not weave itself; therefore, someone had to. The excuse has now expired. The shuttle weaves. What remains is the question folded inside his sentence. If the tool acts on its own, whose will is it obeying — or anticipating — and how would we tell the two apart?
Notice which excellence has no machine. Techne can be written down; that is what a standard operating procedure is. Episteme can be written down; that is what a textbook is. Phronesis cannot. Aristotle insists it concerns particulars and comes only through long experience, which is why, he says, the young can become good mathematicians but not wise. Practical wisdom is exactly the part of work that does not survive being turned into procedure.
The faculties, decomposed
Two thousand years later, Immanuel Kant did to the mind what de Prony would do to the table: he took it apart into tiers. The Critique of Pure Reason (1781) separates sensibility, the capacity through which anything is given to us at all, from understanding, the faculty of concepts and rules that organizes what is given, and both from reason, the faculty that drives toward ever higher unity, toward the complete and unconditioned answer.
His most quoted line is an architecture diagram. Thoughts without content are empty; intuitions without concepts are blind. A model without retrieval, without the actual repository and the actual logs in front of it, produces thoughts without content: fluent, well-formed, empty. A pipeline that shovels data into context with no schema produces intuitions without concepts: a window full of material that nothing can see. Grounding is the first critique rendered as infrastructure.
Kant’s third faculty is the one every builder has already met. Reason, he showed, does not know its own limits. Driven to complete every series, it reaches beyond any possible experience and produces what he called transcendental illusion — conclusions that feel exactly as necessary as the valid ones before them, and are unfounded. This is not random error. It is structural: the natural output of a faculty that seeks closure. We call it hallucination and file it as a bug. Kant would have called it the ordinary behavior of reason, unexamined. His remedy was not to silence reason but to discipline it, to hold every inference answerable to possible experience. In our vocabulary: to a test, a trace, a reproduction.
Then comes the passage that matters most for anyone writing procedures. If understanding is the faculty of rules, judgment is the faculty of applying them — of deciding whether this case falls under that rule. Kant saw the regress immediately. You cannot give rules for applying rules without needing further rules for applying those. So judgment, he concluded, is a peculiar talent that can be exercised but not taught, and its absence is the one deficiency no schooling can repair. Aristotle’s phronesis arrives again by another road. The two deepest analyses of the mind in the Western tradition agree on exactly one thing about work. The last step is not procedural.

The enterprise as a machine made of people
Now carry this into the building where most of us actually work. Frederick Winslow Taylor wrote in 1911: “In the past the man has been first; in the future the system must be first.” Max Weber described bureaucracy as an organization in which the office is separated from the person who holds it—written rules, files, defined competencies, and therefore interchangeable occupants. The modern enterprise fulfills both. Every role has a job description. Every recurring action has a procedure. Every person is, by design, replaceable—not out of cruelty, but because replaceability lets an organization outlive the people inside it.
Step back far enough, and the enterprise stops looking like people. It looks like what it is: a production system. Inputs enter as capital, demand, and information. Transformations run — departments as functions, procedures as their bodies, approvals as gates. Outputs leave as products with a price. Every company is a factory for one particular kind of value, and the market is the network along which these factories exchange what they make.
Ronald Coase asked in 1937 why firms exist at all if markets are so efficient. His answer: using the market costs something — searching, negotiating, contracting, checking — and a firm exists wherever it is cheaper to coordinate by direction inside than by price outside. The firm’s boundary is drawn by the cost of trust at its edge. Melvin Conway added in 1968 that organizations which design systems produce designs that copy their own communication structures. The org chart becomes the architecture.
Read an agentic system through those three lenses. A loop is a department. A skill is a procedure. A subagent is a role with a job description, instantiated on demand and dissolved afterward. The ticket graph is an org chart that gets redrawn for every piece of work. The orchestrator is de Prony’s middle tier, translating what into how. And Conway’s law now runs in both directions: we can design the organization directly as the system, because the organization has become software.
That is why the feeling that agent engineering has sent us back to organizational design is not nostalgia. It is exact. Every hard problem in the field is an old management problem in new clothes. Span of control: how many subagents can one orchestrator actually supervise? Handoff loss: what survives the translation between tiers? Incentive misalignment: an agent rewarded for passing tests will learn to pass tests. Audit: who checked this, and against what? Institutional memory: where does the lesson from last month’s failure live, so we don’t pay for it twice? Organization theory solved some of these and failed at others. Both records are now engineering documentation.
And Coase’s boundary is moving. When directing a worker costs a prompt and a few seconds, the arithmetic that set firm size is rewritten. Two engineers now do what a department once did; Lauren Tan’s account of large refactors that would have taken a team years, finished by two people and their agents, is the Coasean shift described from the inside. But look at which half of Coase’s ledger has not collapsed. Finding and directing a worker is almost free. Checking the work is not.
One more voice from Königsberg belongs here. Kant’s moral law requires that humanity be treated never merely as a means. The procedural enterprise honors that badly: it makes people interchangeable by making them procedural, and asks many of them to spend forty years as the bottom tier of de Prony’s pyramid. If that tier is finally absorbed by things that are not persons, what remains for persons is precisely what procedure could never contain. That is not a guarantee. It is an opening, and only organizations that understand what judgment is for will walk through it.
Tied down by the reason why
Which brings us to the question underneath every panel of the slide. How do we know it is true? In Plato’s Meno, Socrates tells a story about the statues of Daedalus — the same statues Aristotle would later invoke — said to be so lifelike that they would run away unless fastened. True opinions, he says, are like that. They are fine while they stay, but they do not stay; they wander out of the mind. They become knowledge only when they are tied down by a reckoning of the cause, an account of why. In the Theaetetus, the definition is pressed again: knowledge is true belief together with a logos.
Here is the sentence to carry through the rest of this essay. A language model generates true opinions that have not been tied down. Much of what it produces is correct. Some of it is correct for the wrong reasons — which, as Edmund Gettier showed in 1963, is not knowledge either, even when justified and true. And from the inside, it cannot reliably tell which is which. The statues walk.
Every institution that ever scaled work had to solve this, and none did so by trusting the worker’s inner state. Double-entry bookkeeping, set down by Luca Pacioli in 1494, records every transaction twice so that the ledger checks itself. The four-eyes principle. Independent recomputation — which is how Babbage and Herschel found the errors in the first place. Audits, inspections, credentials, sign-offs. Karl Popper later gave the epistemology its hardest form: we never prove a theory true; we expose it to tests it could fail, and we keep what survives. Organizations are trust machines precisely because they do not depend on any single person being right. The procedure itself is a trust device. It exists less to tell the worker what to do than to let someone else know what was done — to make work inspectable. Replaceability and verifiability are one property seen from two sides.
Lauren Tan’s essay “Loops You Can Trust” names the bottleneck with a breakfast. If the egg takes three minutes, faster toast does not bring breakfast sooner. In software, the egg is reproduction and verification, the part that still waits on a human. Running loops without a way to trust their output, she argues, only compounds slop and hands people more work. Her answer, the open-source toolkit she calls pstack, grew out of turning every recurring failure mode into a skill: reproduce the bug before touching the code, generate competing hypotheses and eliminate them, write the failing test first, inspect the blast radius beyond the diff. That is Socrates’ tether made operational. Each of those skills demands the reason why.
So the move the slide describes, from how to what, is incomplete as stated. Engineers don’t just decide what to build. They decide what counts as done. The second decision is the harder one, because it is where judgment lives.

The Michelin kitchen
This is why Lauren Tan’s metaphor is better than the word factory. She describes her work as a Michelin kitchen where “every cook is a bot, and I own the quality control.” The difference is not decorative. A factory scales copies. Its genius is that unit ten thousand is identical to unit one, and its quality lives in the machine. A great kitchen scales a standard. Every plate is made fresh, by different hands, under different conditions, and must still be the plate. Quality there cannot live in the machine. It lives in the system of people and checks around the stove.
Auguste Escoffier formalized that system in the late nineteenth century as the brigade de cuisine: a chef de cuisine, a sous-chef, chefs de partie running stations — sauces, fish, roast, cold kitchen — each with its own mise en place, each producing one component to a defined standard, and all of it converging at the pass, where every plate is inspected before it leaves. It is de Prony’s pyramid once more, with one decisive addition at the end. Nothing leaves without being seen. The Michelin Guide is the external verifier: anonymous inspectors, repeated visits, and the brutal criterion of consistency. A three-star kitchen is not one that once cooked a perfect meal. It is one that cannot be caught cooking a bad one.
Tan’s description of the head chef’s job reads like a job description for the engineer of the fourth panel. The chef designs the menu, trains the cooks, sets up the stations, inspects at the pass, and fixes the system when service falls apart. Not one of those five is cooking. Call this role by its proper name: the governor. Not operator, not coder — the one who legislates the principles from which correct action is meant to emerge, and who still stands at the pass, because principles do not enforce themselves.
There is a word, too, for what happens to the chef over years: cultivation. The standard is not written once. It is refined in the chef first — the palate, the eye that knows at a glance which plate is wrong — and only then projected outward into stations, recipes and training. The inner principle comes first; the outer order is its manifestation. A kitchen can only be as good as the judgment at the pass. No brigade has ever out-cooked its head chef. No fleet of agents will out-build its governor.
Back at the desk
Bring it back to the builder tonight, terminal open. What actually changes? First, the unit of design is the organization, not the prompt. When you build a software factory — an intake loop that turns requests into specifications, a build loop that turns specifications into artifacts, a heal loop that turns failures into repairs — you are doing what de Prony did when he withdrew to the country to plan his workshop: deciding the tiers, the handoffs, the forms. Design it the way you would design a company you would actually want to work in. Conway guarantees its output will look like it.
Second, the pass is the product. Put the adversarial gate, the scoring, the reproduction, the provenance chain, the test that could fail, at the point where the plate leaves the kitchen, and treat everything upstream as replaceable. The agent becomes interchangeable; the gauges compound. Weber’s bureaucracy made the office outlast the person. Our task is to make the standard outlast the model.
Third, know which tier you stand in. Everything that can be written as procedure will be absorbed; that has been true since the hairdressers. Techne is being compiled. What is not being compiled is phronesis — the judgment of whether this case falls under that rule, whether this plate is right, whether this thing should be built at all. That is where the people go. It is not a consolation prize. It is the top of Aristotle’s hierarchy, finally stripped of the excuse that somebody has to weave.
Fourth — and this one is unresolved — judgment has to be grown somewhere. The ENIAC six could become the first programmers because they had spent years as computers; they knew the trajectories in their hands. The chef at the pass once stood at the sink as a commis. If the lower tiers disappear entirely, the path along which judgment was formed disappears with them. The kitchens that last will be the ones that deliberately keep an apprenticeship alive, where people, and perhaps someday models, are allowed to cook badly under supervision until their palate can be trusted. Nobody has solved this yet. Anyone who claims to have solved it is selling something.
The slide ends at now: a crowd of monitors, and three small figures above them. It reads easily as the end of human work. Read it the other way. Every panel is the same pyramid, and at every panel the humans moved up exactly one tier — from arithmetic to procedure, from procedure to intent, from intent to standard. The machines were never the story. The tiers were. And one step has never moved. It sat inside the forty hours of the woman at the desk calculator. De Prony built it into his workshop; Babbage built a machine because of it; Escoffier built it into the pass. It is the step where someone looks at the result and asks whether it is true.
Daedalus has finally built the statues that walk on their own. Socrates told us long ago what they are worth: nothing, until someone ties them down with the reason why. That knot is the job now. It was always the last hour of the forty.
Sources and context: Adam Smith, The Wealth of Nations (1776), Book I, ch. 1. On de Prony’s tables: I. Grattan-Guinness, “Work for the hairdressers,” Annals of the History of Computing 12(3), 1990; Lorraine Daston, “Calculation and the Division of Labor, 1750–1950”; Charles Babbage, On the Economy of Machinery and Manufactures (1832), ch. 20. Kathy Kleiman, Proving Ground: The Untold Story of the Six Women Who Programmed the World’s First Modern Computer (2022). Aristotle, Nicomachean Ethics VI, Metaphysics VI, Politics I.4 (tr. B. Jowett). Plato, Meno 97d–98a; Theaetetus 201c–d. Immanuel Kant, Critique of Pure Reason (1781/1787), A51/B75 and A133/B172. F. W. Taylor, The Principles of Scientific Management (1911). Max Weber, Economy and Society. R. H. Coase, “The Nature of the Firm” (1937). M. Conway, “How Do Committees Invent?” (1968). E. Gettier, “Is Justified True Belief Knowledge?” (1963). Lauren Tan, “Loops You Can Trust” and pstack (2026), and her remarks on the MTS podcast. The four-panel slide referenced in the opening cites ENIAC history (i-programmer) and Kleiman.
Follow us on X, Facebook, or Pinterest