The bottleneck is judgment.
AI is making execution cheap. Building, writing, coding, designing: the machines can do it. The question is no longer who can do the work. It's who can decide what's worth doing.
That takes clarity. It takes the ability to step back, see what you are actually trying to do, and hold that vision steady under pressure. That capacity has a name. Agency: the ability to recognize what you genuinely want and act on it.
If judgment is the only uniquely human role left, stabilizing it is infrastructure.
The biological constraint.
Agency is biological. It runs on specific neural systems that degrade under stress, uncertainty, and overload. The cognitive machinery you need to think clearly about your future is the same machinery that goes offline when you are anxious about your future. Not a metaphor. A well-documented neurobiological constraint.1
Your brain has a circuit that continuously monitors one question: do my actions produce outcomes? When the answer is yes, it suppresses the stress response. You stay capable. When the answer is no, stress deepens.
The problem: that circuit runs on the same hardware that stress already took offline.2 So when someone offers advice while you are overwhelmed, your brain cannot register it as help.
When help backfires.
The clinical record is unambiguous. A well-intentioned intervention delivered to a system that cannot process it does not land neutrally. The exploration of ambivalence with a pre-decisional client measurably lowers commitment to change.3 The advice the friend gave you in the middle of a hard week did not feel like advice. It felt like one more demand.
Most AI mental-health tools — and a fair portion of human-facing systems generally — treat the recipient as a stable processor and the message as the variable. The recipient is not stable. The intervention is the stressor, when it is.
Nearly every AI tool that proposes to help is built on a biologically false assumption: that the person receiving help can process help. The entire approach is structurally broken — not by accident, but by design.
What good assistance should preserve.
A useful system reads the state of the person in front of it, restores the conditions for clear judgment, and stays out of the way of what that person actually wants. The human controls the ends; the machine observes, estimates, and stabilizes. It never decides what you should want.4
Eyeglasses versus GPS. Glasses correct a structural limit and expand your capability. They do not make your eyes weaker. GPS provides efficiency but degrades your internal navigation over time. After a decade of GPS, your phone dies and you are lost.
The success metric: your capacity for self-directed action is higher after using the system, even once it's removed. You keep it because it extends you. If you are worse without it, it failed. It became GPS.
The system is governed by constraints that make it structurally impossible to optimize for engagement instead of capacity. As an engineering requirement, not a policy.
The notebook is the work.
The substance is the notebook. Axioms, constructs, mechanisms, predictions, kill conditions, open problems. Every claim graded by epistemic status.
It is open for critique because critique is the only way to find out whether the specification survives contact with people who know more than I do.
The human always decides where to go. The system makes it possible to see clearly enough to decide.