Identity by Design? Evaluating Gender Conditioning in LLM-Generated Agent Identity Profile

In multi-agent reasoning frameworks powered by large language models, agent roles are often instantiated through identity descriptions that condition their behavior. This paper investigates whether and how the gender assigned to the agent responsible for defining role-specific identity profiles affects the linguistic identity, sentiment, and gender expression of downstream agents. We introduce an extensive corpus of agent identity descriptions generated under controlled combinations of frameworks, roles, models, and gender conditions. Through quantitative and qualitative linguistic analysis, we observe a consistent skew toward female identity across models and roles when gender is unspecified, along with varying degrees of polarity and subjectivity depending on the description framework. Notably, cognitively-oriented frameworks suppress affective expression, while trait-based frameworks amplify gender alignment. These results reveal that identity conditioning is not solely determined by prompt parameters, but emerges through a layered interaction of model priors, framework semantics, and role-specific expressive constraints.