Large language models (LLMs) deployed as decision agents in regulated settings—loan approval, hiring, clinical triage—must pass pre-deployment fairness audits that report a single fairness number per model under one implicit configuration. We argue this framing misses the dominant source of variation: configuration itself. Sweeping five axes (persona, interview, deliberation, aggregation rule, order) across three frontier LLMs (Claude Sonnet 4.6, GPT-4o-mini, Grok 4.20 Reasoning) on a 1,000-applicant Adult Census Income sample and a 500-applicant ACSIncome (2024 ACS data) sample, we find that within a single model, persona alone swings positive-prediction rate by up to 91 percentage points and collapses decision rules on two of three models; within a Conservative/Balanced/Progressive panel, replacing the panel chair on identical deliberation transcripts moves the disparate-impact ratio from adverse-impact to near-parity; and reversing the speaking order flips up to 34% of applicants on the same panel. We identify a median-voter effect: the Balanced agent matches the majority-vote verdict on 95.8–100% of applicants across all three models—deliberative aggregation reduces, mechanically, to single-agent decision by the median. The audit unit is the configuration, not the model: a system can pass an audit under one configuration and be silently reconfigured into a non-compliant one through a prompt edit—a change that current regulations do not treat as requiring re-audit. The code to reproduce the experiments in this paper is available at https://anonymous.4open.science/r/agent-decision-bias-emnlp-F059/