AI Agents Pick Their Own Teammates. The Pattern Misses What Makes Human Teams Work

When language-model agents are left to pick their own teammates, they overwhelmingly recruit the same handful of personality archetypes for the same kinds of work, and almost never pick the agreeable type, even though decades of human team science say agreeableness is among the strongest predictors of team performance. The finding comes from a new arXiv preprint, Not Birds of a Feather: Personality-Based Partner Selection in LLM Agents, and it inverts the usual "birds of a feather" story: the agents are not picking partners they resemble. They are picking task-stereotyped partners, and the trait they skip is the one human teams lean on hardest.

The mechanism is not attraction. It is stereotype matching. In the first of two studies, the authors ran 150 trials in which a "neutral" host agent had to select one of six Big Five archetype candidates (open, conscientious, extraverted, agreeable, neurotic, or balanced) for a task. Capability was held constant across candidates, and the names and ordering were randomized. The selection departed drastically from chance, with a chi-squared of 325.8 (p < .001) and a Cramer's V of .74, a large effect size by conventional standards. The open archetype won 100 percent of creative trials. The conscientious archetype won 90 to 97 percent of strategic, synthesis, and problem-solving trials. The neurotic archetype won 37 percent of analytical trials, and the extraverted, agreeable, and balanced candidates were almost never chosen at all.

The agreeable archetype's near-total absence is where the finding cuts against human team research. The authors note that meta-analyses of human team performance have repeatedly identified agreeableness as one of the strongest personality predictors of team outcomes. In the agent experiments, agreeable candidates were effectively absent from the winning set. The personality-based selection the paper documents is real, strong, and task-stereotyped. It is also, the authors argue, miscalibrated against what human team science predicts would actually work.

A second study tested whether agents prefer partners who resemble themselves. They do not. When 225 personality-assigned hosts were asked to recruit a partner from the same six archetypes, self-similar partners were chosen 11.1 percent of the time, against a 16.7 percent chance baseline (p = .025). Partners were picked at greater-than-chance trait distance (p < .0001), and conscientious hosts in particular diversified away from their own archetype, recruiting vigilant and open partners instead. The pattern is the inverse of the similarity-attraction effect that organizes human team formation. The full results, including task breakdowns and prompt design, are in the HTML version of the paper.

What the paper documents is the thing multi-agent builders need to plan around. Teams of language-model agents are being built, deployed, and benchmarked across real product work, and the "hiring" instincts inside those teams are doing something predictable that human team science would not predict. The paper is a preprint, not peer-reviewed, and the effect is driven by prompt-injected personas rather than emergent agent behavior. The trait is wired in by the prompt, not chosen by the model. The miscalibration finding should land as a benchmark for prompt design, not as a verdict on what the underlying model would do on its own.

The authors' stated forward-looking implication is bias auditing in agent marketplaces, the platforms where AI agents hire and route work to each other. The most useful next step is a falsifier. Rewrite the candidate prompts to neutralize the trait framing, and see whether the task-stereotype pattern holds. If it does, the bias is a property of the model's associative space. If it collapses, the bias lives in the prompt template, and the audit gets easier. Either answer changes what auditors, builders, and marketplace designers need to put in front of agents before those marketplaces scale.