I want to present a simple thought experiment.  Imagine a pull request comes into your queue. Nothing exotic, the kind of change that ships a hundred times a week on most engineering teams.  At this point, most of us may dispatch an AI review before we even look at it. I have been working on an experimental swarm of AI agents who I send out to review the request before I look at it or it merges. They use the same underlying models in every one of them with the same code in front of all of them. The only thing that differs is how each agent's identity is conditioned.  (How they are conditioned is wildly interesting, but that’s a topic for a different post) Most of them clear it, but one stops on a single line, a boundary case the others read right past. For anyone that has been in software development, it’s the kind of subtle miss that pages someone at 2am three weeks later. Once I had collected enough of these cases to measure real trends, I looked at the data and noticed about 58% of the issues the swarm catches are found by exactly one agent. One agent sees it and every other agent, looking at the identical code, walks right by. So if I'd run any single one of them alone, even a strong one, I'd have approved more than half of what the group caught. Same model. Same code. The difference was who I asked it to be.  I have been talking a lot about the role of identity conditioning in model behavior, and have written about it in more detail here. xitter.cf/gabaum10/status/207666…