AI agents picking teammates default to personality stereotypes

Researchers ran 375 trials where LLM host agents picked collaborators from six Big Five personality archetypes, with capability held explicitly constant across every one. Human meta-analyses rank agreeableness among the strongest predictors of team performance. The agents did the opposite anyway. They had no capability difference to justify the call, per the paper (arXiv 2607.19785).

The results weren't close. The open archetype won 100% of creative trials. Conscientious won 90 to 97% of strategic, synthesis, and problem-solving trials. Agreeable and extraverted archetypes were almost never picked. The statistics back up how far this drifted from chance: χ²(5)=325.8, p<.001, Cramér's V=.74.

There's a second finding that's more useful for builders. When hosts were themselves assigned a personality, they chose self-similar partners below chance, not above it. That cuts against the assumption that personality-matched agents cluster together.

The paper doesn't say whether this bias shows up in production systems that route tasks by persona tags. It also doesn't say whether the skew is an artifact of how these six archetypes were described to the host agent. That gap matters for anyone designing a marketplace that exposes personality or tone metadata as a selection input.

This is a testable claim: audit your routing logs for the same skew before assuming it's not there. A χ² test on your own selection data is cheap. Finding out from a user complaint is not.