What's in a Name? Addressability, Parental Framing and Group Labels in Generational Populations of LLM Agents

Names in multi-agent simulations of language-model agents are usually treated as scaffolding. We varied them as the experimental manipulation in a generational population of LLM agents (Claude Haiku 5.5) that forage, hunt in pairs, form couples, endow their children, compete with them for survival, and vote on their own scoring rule. Across four pre-registered studies (480 analysed eight-generation runs, about 367,000 model calls), every arm saw identical information and identical random draws; only the handle each agent carried, and how that handle was described, changed. Any stable individual handle, whether a name or a bare ID and whoever gave it, raised pair hunting two- to fivefold over no handle. Couple formation instead followed a parental frame: a meaningless ID raised it when parents picked it, and equally when agents were merely told their parents had given it. Only names composed by the parents raised investment in newborns, and hunters with such names shared their kills most often. Arbitrary Blue/Green bands produced in-group pairing and hunting, yet a band added to an individual ID doubled pair hunting rather than splitting agents into factions. Set against human findings registered in advance, identification and group-label effects matched in direction, but namesaking did not raise parental investment, and agents put about 3% of their holdings into a public good, against the 40-60% typical of people. Exploratory analyses suggest that a world-assigned ID lowered couple formation below no handle at all, that parents gave less to a child the lot placed in the other band, and that a single sentence saying children are named after a parent halved couple formation. How agents are named, and how naming is described, shapes cooperation and family formation in LLM agent populations.