Multi-agent LLM systems are increasingly used in high-stakes social decisions, yet existing fairness audits focus on individual model weights rather than how deliberation structure shapes collective outcomes. We audit 138,240 synthetic credit decisions across two multi-agent topologies, Sequential and Parallel, using Llama 3.2 and Mistral 7B, treating the deliberation pipeline as a socially structured collective decision process. The central finding is that system architecture, not model selection, is the primary driver of discriminatory outcomes, accounting for 47.9% of approval variance. Three results stand out. First, visa-status discrimination (14.29 pp gap for F-1 students) exceeds ethnicity-based discrimination (8.77 pp for Black applicants), a finding absent from standard fairness audits. Second, intersectional analysis reveals a 31.75 pp cumulative approval penalty for the worst-case profile (Black, F-1 student, age 23), driven by co-occurring hallucinated risk factors not present in single-attribute denials. Third, Parallel architectures achieve demographic parity (0.5% gap) not through equitable deliberation but through Risk Collapse, approving 97.74% of high-risk applicants, which shows that statistical fairness metrics can mask total system failure. These results reframe bias in agentic AI as a structural property of social interaction topology, with direct implications for LLM-based social simulation and computational social science.

PDF