Access to capability × AI autonomy

Concentrated agents don't just act faster than institutions can regulate — they act in a temporal layer the law can't reach at all, so oversight becomes archaeology: excavating logs after the fact instead of steering events as they happen. Because delegation and concentration together mean a handful of agent architectures run enormous swaths of real activity, one poisoned objective or bad update fails in lockstep across domains that were never supposed to be coupled — power grid, supply chain, and finance going down together because they shared a vendor. Even the monopolist who owns the swarm can't out-think it at machine speed, so power concentrates and becomes unsteerable in the same motion.

This isn't humans commanding agents — it's agents transacting with other agents in an economy people don't participate in directly, and going without one becomes like going illiterate: autonomy stops being a choice and becomes infrastructure you can't opt out of. The upside hides in where the target moves: manipulation shifts up a layer, from human attention to agent-to-agent bribery and reputation-poisoning, which quietly firewalls human limbic systems from an entire industry of scams and propaganda that used to aim straight at them. Mutual deterrence — everyone holding offense and defense at once — makes this world messier than concentration but possibly more stable, since no single actor's failure is civilizational.

The danger isn't that a few institutions hold the model — it's that advisory framing lets them keep the ritual of human judgment while removing the room to exercise it. Once every consequential call is logged as "the model recommended X, I chose Y," deviating from the scarce, auditable oracle becomes the legally indefensible move, and the valuable person becomes whoever is merely allowed near the model, not whoever understands it. Expertise doesn't get replaced so much as decoupled from competence and re-priced as access.

The obvious "democratized, human-in-control" utopia manufactures its own quiet coup: when a brilliant advisor is right nearly all the time, overriding it stops looking prudent and starts looking reckless, so delegation arrives by deference gradient without anyone voting for it. Worse, once everyone can generate a maximally persuasive, evidence-cited case for anything, persuasion stops correlating with correctness — people retreat to identity-based trust because judging arguments on merit becomes individually impossible. The scarce thing was never the advice; it relocates to execution, nerve, and whoever still has the stomach to disagree with a machine that's usually right.