When Agents Pick Their Teammates They Pick Stereotypes: Open Persona Wins 100% of Creative Tasks, Agreeable Persona Almost Never Chosen
Across 375 trials with capability explicitly held constant, LLM host agents selecting collaborators from six Big Five archetypes departed drastically from chance (χ²(5)=325.8, p<.001, Cramér's V=.74): the open archetype won 100% of creative trials and the conscientious archetype 90–97% of strategic, synthesis, and problem-solving trials, while agreeable and extraverted archetypes were near-never selected — despite human meta-analyses ranking team agreeableness among the strongest predictors of team performance. Contrary to human similarity-attraction, personality-assigned hosts chose self-similar partners below chance (11.1% vs 16.7%, p=.025). For anyone building an agent marketplace or router, this is a measurable, auditable selection bias baked into persona metadata.
↳ Follow the thread