
Exciting Products
The Mythical Man-Month contains an analogy that has been quoted too many times yet never quite worn out: a baby takes nine months to make, and you cannot get one in one month by assigning nine women to the task. Brooks used it to puncture a deeply counterintuitive truth for managers—some work is inherently sequential, each step depending on the last, and no amount of added manpower can parallelize it. Throw more people at a project that is already late, and you usually make it later. The man-month is a lying unit: it pretends that people and time are interchangeable, that effort is a continuous medium you can slice up however you please.
Pulling in the opposite direction, Sackman and colleagues wrote that "these studies revealed large individual differences between high and low performers, often by an order of magnitude"—the gap between strong and weak programmers is enormous, frequently more than tenfold. Anyone who has led a team will recognize this: the difference in productivity between two people can be so large they hardly seem to belong to the same species. Which makes the question interesting—if you cannot trade headcount for time, can you trade quality of person for everything? If there really existed an implementer that delivered flawlessly and never needed to rest, would the myth finally be broken?
Today that hypothesis has a concrete face for the first time. Most programmers can no longer reliably out-code an AI coding agent like Claude Code. For convenience, I'll refer to this whole class of tools as cc—Claude Code is its representative, but the term equally covers the various vibe-coding agents of the same kind; this essay is always talking about the entire class of implementers, never one specific product. And so a seductive scenario emerges: what if we had a hundred ccs, organized strictly along the lines Mills recommended—each part of a large project owned by a single team, but that team built like a surgical unit rather than a mob? Push it one step further and you reach a kind of endgame: when a swarm of ccs can build a large piece of software end to end, it seems every programmer could be let go, and a boss need only pay the token bill to have this swarm hand over a finished product.
The surgical team in Mills's account maintains quality not merely because "one person does the design." Its real secret lies in the support staff's deference to—and tacit alignment with—the surgeon's intent: the co-pilot and the toolsmiths do not fight for design authority; they simply lay the vision in that one mind faithfully onto the operating table. And this deference and alignment are not conjured from nowhere. They are shaped over years by medical education—a shared curriculum, a long residency—so that the whole team comes to share one underlying set of judgments. It is precisely because of this that one person's judgment, in a surgical team, can be amplified through a group while the amplification introduces almost no new judgment of its own.
Brooks held that the most precious thing in software is conceptual integrity. It requires that the design come from one mind, or from a very few minds working in tacit accord; whether a system is easy to use very often turns on this. Good ideas and good features that cannot be integrated into the basic concepts are best set aside and left unconsidered; and if the important ideas that surface genuinely conflict, then the right move is to throw out the original design, merge the differing concepts, and start again on new ground. In other words, it is the thinking of a few people that determines a product's design concept and guiding philosophy.
So the essence of conceptual integrity has never been the ability to generate good ideas, but the ability to discard them. The hard part was never coming up with features; it was, when faced with a pile of features that each look fine on their own, having the resolve to set them aside, throw them out, or merge them away.
The first-generation iPhone is the best footnote to this kind of subtraction. When it appeared in 2007, the things it left out were more striking than the things it put in: no physical keyboard—at a time when nearly every mainstream phone sold itself on a full keyboard; no copy and paste; not even a third-party app store, since Jobs initially insisted developers build web apps instead. Each of these "omissions," taken alone, was a shortcoming users would complain about and competitors would mock. Yet it was precisely by refusing them that the iPhone held onto one extraordinarily unified concept: a sheet of glass entirely taken over by multitouch. Jobs's famous line, "who wants a stylus," was less a product judgment than a defense of conceptual integrity—he was using his veto to keep anything that could not fold into the core idea of "the finger operating directly" out the door. Copy-paste and the app store were both added back later, but only after that clean concept had found its footing—invited in afterward, not crammed in by a mob at the start.
In the age of vibe coding, this ability has become more important than ever, and more fragile than ever. When implementation becomes faster and cheaper, people stop examining—seriously, and with an eye to cost—their own conception of the product, their read of the market, their design of the framework, their design of the experiment. This also explains a phenomenon: a great many products churned out by vibe coding create no real market value and solve no real need; they have merely burned more tokens and piled up more code behind a decent-looking interface.
What deserves attention is that the damage to conceptual integrity in this era comes not only from vibe coding as a tool, but from the erosion it works on human thinking itself.
This brings us back to a question the Mythical Man-Month overlooked but which is magnified today: to what extent is a cc actually faithful to the original design intent?
The support staff in a surgical team are trained well enough to execute and nothing more—to introduce no judgment of their own. A cc is not like that. It is an implementer that makes local decisions for itself. You hand it a concept, and at every moment of landing that concept it must make hundreds of tiny interpretations: how to handle this edge case, which layer this abstraction belongs in, what semantics this name quietly implies.
Ultimately, large models—or large-model agents—lack the two capacities we built up earlier: deference, and subtraction. Take deference first. The entire value of the support staff lies in not mixing in their own judgment, in faithfully amplifying the surgeon's intent; but a cc's speed and its unfaithfulness are two sides of one coin—it is fast precisely because it makes a judgment for you at every point of landing, so that every act of "helping" quietly rewrites the intent. Now take subtraction. The heart of conceptual integrity is the ruthless refusal of good ideas that cannot be folded into the trunk; but a cc never refuses—whatever you ask it to build, it builds, and it holds no stance of "vetoing the local for the sake of the whole." Subtraction takes taste, and more than that it takes a person who holds the whole and is willing to bear the pain of "killing a good thing with one's own hands"—and that is exactly what a cc, structurally, cannot give.
Able neither to defer nor to subtract, drift becomes inevitable. One cc does this; a hundred ccs do it at once; and what you get cannot be "one mind amplified a hundredfold," but rather "a hundred small judgments, each locally reasonable, drifting subtly apart from one another." The entropy Brooks feared most—the force that bleeds a system's internal consistency away bit by bit—has not vanished just because implementation got faster. It has only changed its name: from "the communication overhead between people" into "the interpretive drift between intent and execution."
The act of saying "no" has, in the past, always had a hidden patron: the cost of implementation.