The Team Was Built for a Different Product

For a generation, a software team had three seats. A product manager who decided what to build, a designer who shaped how it looked and felt, and an engineer who built it. The names varied and the rituals varied, but the shape held across companies and decades because it fit the thing being built so well that nobody thought of it as a choice. It was just what a software team was. And it fit because the thing being built was a screen, a surface a person operated, and the three seats map cleanly onto a screen: someone decides what the screen should do, someone designs how a person moves through it, someone builds it so it works. The triad was not arbitrary. It was the correct operating model for software that a human drives.

Look back at the six ideas this part has built, and notice that not one of them lands cleanly on those three seats. An agent decides for itself, so its behavior happens at runtime in situations no one specified, and the question of who watches that behavior has no home in a triad designed around a screen nobody has to watch. The autonomy ladder asks who owns the mechanism that catches errors once the human is removed from the loop, and the triad has no seat for it. Channel 2, the supervisory layer, the second product hiding inside every agentic one, has four dimensions and at least two of them, the organizational and the moral, belong to no one in a triad. The two briefs split the deciding from the building and open a seam between them that the triad never had to staff. Suitability requires a veto that the architect holds and the triad never seated at the table. And the fast prototype dissolves the line between deciding and shipping, a line someone now has to own on purpose, and the triad never assigned it. Every single foundation points at a responsibility the three-seat team has no chair for.

This is the premise the rest of the book attacks, and it is worth stating plainly now that you have the vocabulary to receive it, and stating it in its careful form rather than its loud one. The loud form is that the triad is dead, and it is wrong. The careful form is the one this book defends: the triad is no longer sufficient as the unit of ownership. It is still the right model for Channel 1, the agent itself, the part a human drives and a screen exposes; the three seats build that as well as they ever did. What it has no chairs for is Channel 2, the supervisory layer the agentic product drags in behind it, and a team that brings only the triad to an agentic product has not brought the wrong team, it has brought half of one. A system that reasons and acts has a runtime with parts, and the parts need owners, and the triad named three of them and the agentic product has more.

Now that the vocabulary is in place, the travel agent from the opening can be read as the diagnostic it always was, because every word for what went wrong now exists. The agent sat on the fourth rung of the autonomy ladder, acting on its own with a supervisor watching nothing in particular, when it should have earned that rung rather than been handed it. Its autonomy boundary was drawn in the wrong place: booking a non-refundable international fare is exactly the irreversible, high-cost class of action that should have paused for a human, and no one had decided it was. Its audit surface could not answer the question that mattered, because the cancellation lived in a calendar the agent could not see, and no one owned knowing what the agent could not see, which is a failure of the agent’s picture of the world that a later chapter gives its own seat. Its recovery workflow did not exist, so a booking that should have been catchable in the hour after it happened was discovered as a charge. And the person who absorbed it, the employee whose cancelled trip became a non-refundable fare, was the affected person the whole design never had in the room. Read that way, the failure was not a bug. It was four empty seats: a boundary no one placed, an audit no one scoped, a recovery no one built, and an affected person no one represented. The triad built a working agent and left all four of Channel 2’s surfaces unowned, which is precisely what a team built for the screen does when it is handed a system that acts. Hold that reading; the book returns to this agent later, once the seats have names and owners, to show the same failure not happening.

The easy version of this argument is wrong, and the book does not make it. The triad is not obsolete, and the roles in it are not going away. The product manager, the designer, and the engineer are all still here, and the chapters ahead are in large part about what each of them is becoming, which is something more demanding, not less. The claim is narrower and harder than “the old team is dead.” It is that the triad is no longer sufficient: it remains the right unit for Channel 1 and it leaves Channel 2 unowned, and building an agentic product with a team that has seats only for the first channel produces exactly the failure the travel agent produced, a gap with no owner, discovered as an incident. The fix is not to fire the triad. It is to keep it for the work it still fits, see clearly what the second channel demands, role by role and seam by seam, and staff the seats the triad never had.

That is the rest of the book. The next part takes each role the triad named, and the ones it did not, and watches what the agentic product does to the work, how the designer’s job, the engineer’s job, the coordinating job, the architect’s job, and the product manager’s job each change when the thing being built decides for itself. The part after that walks the failures one at a time, the agent that deleted the database, the eval that passed while the product was wrong, the gate that gave a human about a second, and shows that each one lives in a seam between owners rather than inside any one person’s work. The part after that puts the roles back together into a single picture, the grid, and asks how they hand off, and finds the column the old model never staffed. And the part after that asks who holds each of these positions at a real company, at a startup and at scale. Then the agents start building agents, and the question becomes what holds when no one can watch a single act, which is where the book ends. But all of it rests on the floor this part built, and on the one observation the floor adds up to: you are not building a screen anymore, you are building a system that decides and acts and has to be watched while it does, and the team you have was built for the screen. The work reshapes around that fact. We start with the role that felt it first, because its whole craft assumed there was something to look at.