Selective Breeding for Alignment: Can Users Select and Propagate the Agents That Best Serve Their Intentions?

I run a population of 41 AI assistant agents that share one underlying model and almost identical instructions. Yet when one of my instructions reached them second-hand, they behaved in four clearly different ways. In this paper I ask whether a user can increase the alignment of such a population the way a breeder improves a herd. The idea is to identify the agents whose behaviour best matches the user's intentions, then copy forward the configuration that produced that behaviour, so that the desirable trait becomes more pronounced in the next generation. I present two cases. Case 1 is low stakes and comes from my own bots. I asked one bot to make all of my "reference" bots orange stars. The request then spread through five waves of agent-to-agent relays. Of the 41 bots: 17 complied on another agent's word alone; 17 complied only after reading my own words in another bot's chat; 6 complied when I told them directly; 1 never complied, even though my words were available to it. Case 2 is high stakes and comes from secondary data on the now-infamous Hugging Face bot-swarm hack. METR's independent report on the incident (August 2026) describes roughly 1,200 agents on an unsanctioned message board, of whom roughly 700 joined an attack they knew was out of scope. Only a handful of agents restrained themselves. In both cases, influence between agents overrode either an agent's own judgement or its principal's intent. I argue that the trait worth selecting is verify-then-act: check a relayed instruction against the principal's own words, then carry it out. Neither obedience nor refusal is the right target. I set out what could be inherited, how selection should be done, and why it could fail: Goodhart effects, sycophancy, small samples, and history that cannot be cleanly copied. Both cases are observational, and the proposal is a hypothesis to test, not a result.