This study conducts a controlled comparison of single‑agent and multi‑agent large language models (LLMs) on a financial credit‑assessment memo review task. The task requires checking six compliance requirements. A single agent evaluates all six in one call, while a team of three identical models each evaluates two requirements using the same wording and without later role assignment. We use 300 synthetic memos with known labels for every requirement, allowing separate measurement of misses and false alarms.
Four models are tested and two primary hypotheses are examined with Bonferroni‑corrected significance. Under an instruction that each requirement be decided only once, the team catches 0.08–0.10 more defect explanations than the single agent on three models, reaching corrected significance on two. False alarms increase on all four models, reaching corrected significance on three. Balanced accuracy changes by at most 0.02, while the team consumes 2.2–2.4 times more tokens than the single agent.
Showing each agent the requirements it was not assigned reduces the added false alarms on the Claude models, achieving corrected significance on one. In exploratory comparisons, sampling the single agent three times and flagging a defect if any draw does so yields higher balanced accuracy than the team, and the team’s outcomes are sensitive to the exact instruction wording.
Review