Judgement does not reproduce itself

An AI agent ships a flawless code change. The tests pass. The linter is green. Coverage holds.

An engineer blocks it anyway.

He saw what no check measures: this solution moves a responsibility to the wrong place, and it will cost dearly in six months. No machine settled that.

This piece is about that act. About what you have to have lived through to be capable of it. And about the question nobody asks any more: who, ten years from now, will still know how to do it?

That AI writes code, fast and for almost nothing, is a settled debate. What remains is what it cannot do — judge. And the way an organisation keeps making, or stops making, people capable of doing it.

In my two previous pieces, I laid out a thesis I still believe holds. AI has flipped the roles: the machine produces, the human judges. And the sentence that carries all the rest is that judgement sits downstream of craft.

You cannot assess what comes out of a black box without already knowing what good looks like. That knowledge is not downloadable. It is built in contact with the code, line by line, over years.

I stopped there. That was too soon. If I take my own sentence seriously, it turns against my conclusion.

The engineer who blocked that merge could do it because someone paid him, for ten years, to write lines, debug flaky tests and read other people's code. His judgement lives off capital accumulated yesterday.

The junior joining his team today will never build that capital. Look at her day: she reviews generated code, she approves suggestions, she orchestrates. Execution has become abundant and nearly free, so nobody is paid to do it any more.

Yet doing it is precisely what produces, years later, the ability to block the change every indicator waved through.

A job aimed solely at judgement is therefore not a stable destination. It is a livable position for those who reach it from above, with ten years of craft behind them. And a closed one for those who should be joining it from below.

The real question, the one left standing once you have exhausted the individual strategies, is a question of reproduction. How does an organisation keep making people capable of judging? And who controls that mechanism?

It used to be that this knowledge circulated freely between practitioners. A junior learned by reading senior code, by getting corrected in review, by absorbing years of field experience in direct contact with others.

Today, when that same beginner has a question, she asks a model. The learning that passed from an old hand to a new one becomes a service rented from a third party and billed by the token.

This is not a side effect. It is the breaking point. By settling between the old hand and the beginner, the model does not only replace the production of code. It replaces the transmission of the craft, and it charges for it.

The economics need human judgement. Only it arbitrates ambiguity, knowingly lets a risk through, or blocks work that passes every test but leaves you uneasy. The pipeline cannot do that.

But that judgement is expensive, and its price is invisible. Between two competing teams, the one that maintains its competence and the one that lets it melt away ship the same green code. The second ships cheaper. So it wins. Until the day of the accident — the one only the first knew how to block.

What makes this mechanic relentless is that nobody in it chooses atrophy. The executive who understands all of this still suffers it. Every actor makes the rational choice at their own scale, and it is the aggregate that destroys what all of them depend on. The system needs something it destroys at the same time as it exploits it.

A contradiction of that kind is not a fate. It is a point of tension. So it is ground where choices remain possible. Including executive ones.

I built a harness, Ferry, in a context where AI agents produce code continuously. Concretely, it is a series of automated checks: every change, whether it comes from a human or an agent, has to clear them before it earns the merge. In parallel I theorised a role, the Spec Steward — the person who holds the conditions under which a solution has the right to exist in the system.

I had thought of them as neutral technical objects. They are not.

The same harness can serve two opposite projects. An instrument of surveillance that enforces compliance while the human is pushed out of production. Or a control point through which engineers keep their grip on what the machine ships. What decides between the two is not the technology. It is who owns the harness, and who owns what it validates.

There is an irony there that I had noted without weighing it. I built, with AI, the tool that governs the code written by AI. The worker erecting with his own hands the apparatus of his own supervision. But the same apparatus, if it stays in his hands, becomes the apparatus of his autonomy. Everything hangs on ownership.

And this is no longer an isolated trajectory. The creator of Claude Code says he no longer prompts: his work, now, is to write loops, systems where one agent produces while another verifies, and where the human designs the chain rather than operating inside it. The industry has just given that pattern a name, loop engineering. Anthropic, by its own numbers, has Claude write more than 80% of its code, and names what is left to humans: judgement.

So everyone is building harnesses, right up to the model builders themselves. The question almost nobody asks while building them is the one that will decide everything: who will own them.

The engineer holds two roles nobody else can hold. Today, he is the one who settles what the machine cannot. Tomorrow, it is through him or through no one that someone else will learn to do it.

An organisation can do without one of its engineers. It cannot do without all of them and keep both roles, and it will find they cannot be bought back. That is what makes this position, held collectively, a chokepoint. And its individual version a mere reprieve.

The logical next step, then, is not to take refuge in judgement while waiting for the capital to run out. It is to own, wherever you can, what lets judgement reproduce itself.

There is an immediately actionable version, and I address it to an executive. In-house competence kept alive is not a cost to optimise. It is strategic insurance. The day the risk arrives that only human judgement knew how to block, you discover you sold the one thing that was not automatable — and that you cannot buy it back at the price you let it go for.

Maintaining that competence when it costs more than the machine is exactly what you expect from insurance. You pay for the day you will need it, not for the days when everything is fine.

Concretely: keep paying people to go through the step you are tempted to delete. Organise the transmission instead of outsourcing it to the model. Keep ownership of the artefacts that govern production rather than renting it.

None of this is inevitable. Somewhere, today, an engineer is still blocking a change every indicator waved through.

The question of the coming years is not whether the machine will produce more. It has already answered that. It is who will pay so that, in ten years, someone still knows how to perform that act.

P.S. This piece was written with the assistance of an AI. It produced, I judged. How much longer I will still know how to do that is the open question.