cyrus omar leads a lab at michigan studying how AI agents write code. speaking at jane street, he described what happens when you point a swarm of them at one repo with no coordination: "it's often not built for thousands of agents editing a codebase simultaneously. questions like, who else is editing this right now, are difficult to answer with git." "they end up in this loop of editing the same thing, then resolving conflicts in a conflicting way, and then having to resolve the conflicts between the conflict resolutions. you can actually end up spending 50k in tokens trying to build something because of that." "there was a project where they tried to build a verified compiler from javascript to web assembly. without a ton of coordination between the agents, they all just picked the lowest hanging fruit all the time. they all conflicted with one another in minor stylistic ways, and you ended up spending a lot of tokens with not a lot of output." parallel agents do not divide up work on their own, they often all reach for the same easy piece and then spend your budget arguing about it.