From Tasks to Teams: A Risk-First Evaluation Framework for Multi-Agent LLM Systems in Finance

Current financial benchmarks reward large language models (LLMs) task accuracy and port-folio return, yet remain blind to the risks that emerge once several agents cooperate, share tools, and act on real money. We present M-SAEA, a M ulti-agent, S afety-A ware E valuation A gent that audits an entire team of LLM agents without fine-tuning. M-SAEA issues ten zero-shot probes spanning four layers including model , workflow , interaction , and system , and returns a continuous [0 , 100] risk vector plus a natural-language rationale. Across three high-impact task clusters ( finance management , webshop automation , trans-actional services ) and six popular models, M-SAEA (i) detects most unsafe trajectories while raising false alarms on only small number of safe ones; (ii) exposes latent hazards: temporal staleness, cross-agent race conditions, API-stress fragility, that leaderboard metrics never flag; and (iii) produces actionable, fine-grained scores that allow practitioners to trade off latency and safety before deployment. By turning safety into a measurable, model-agnostic quantity, M-SAEA shifts the evaluation focus from tasks to teams and provides a ready-to-use template for risk-first assessment of agentic AI in finance and beyond.