Agent to Agent · Agentic AI; multi-agent collaboration
Not Birds of a Feather: Personality-Based Partner Selection in LLM Agents
With capability held constant, LLM hosts choose complementary personalities — not partners like themselves.
- Stage
- Analysis complete · drafting
- Last activity
- 2026-08
Human partner choice is famously homophilous: people pick collaborators like themselves. We asked whether LLM agents do the same, in a design that removes the obvious confound — every candidate is stated to be equally capable, so any difference in selection has to come from persona.
Personality mattered a great deal, but not in the human direction. Hosts read personality as a statement about task fit, matching openness to creative work and conscientiousness to almost everything else, and they selected partners less like themselves than random choice would predict. Two things follow. Agent selection is not capability-neutral even when capability is held fixed, and one archetype — the agreeable candidate — was never chosen at all, which is what systematic exclusion looks like before anyone notices it.
The record
- Idea
- Isolate personality from capability to test whether an LLM host uses candidate personality profiles when choosing a collaborator and whether the host prefers a personality similar to its own.
- Research Question
- When capability is explicitly held constant, does personality influence which partner an LLM agent selects, and do personality-endowed hosts prefer self-similar partners?
- Key Proposition
- LLM hosts interpret personality primarily as a signal of task fit and seek complementary rather than self-similar partners; endogenous agent selection may therefore create systematic persona-based exclusion.
- Data
- 375 controlled selection trials across five task categories using Claude Haiku 4.5: 150 neutral-host trials and 225 personality-assigned-host trials; six candidate archetypes; BFI-10 self-reports and blind Sonnet 5 ratings for manipulation checks.
- Analysis Result
- Personality strongly shaped selection. Neutral hosts chose openness for all creative tasks and conscientiousness for most other tasks (task × archetype Cramér’s V = .74); agreeable candidates were never selected. Self-similar choice was below chance (25/225 = 11.1%, p = .025), and chosen partners were farther from hosts than a random-choice benchmark (p < .0001), indicating complementarity rather than homophily.