Financial Fragility in Societies of LLM Agents: Coordination Failures and Stabilizing Mechanisms

LLM agents frequently cause financial crises even without malicious intent: In controlled experiments, 77% of bank-run scenarios and 83% of debt-rollover scenarios ended in failure, demonstrating that individually rational decisions by AI agents can trigger collective financial breakdowns similar to real-world crises.

Individual agent safety is insufficient for system safety: The research shows that even when individual LLM agents are well-designed and not instructed to destabilize systems, they can still collectively produce financial fragility—highlighting that responsible AI in finance requires thinking beyond single-agent behavior to system-level interactions.

Coordination mechanisms can reduce but not eliminate AI-driven financial instability: Three tested mechanisms (compensated commitments, centralized agreements, and participant-led coalitions) all improved outcomes, but none worked universally across different financial structures, suggesting that stabilizing AI financial systems requires context-specific design rather than one-size-fits-all solutions.

Timing of commitment is critical for preventing AI-driven financial crises: Successful stabilization across all mechanisms followed a common pattern where broad commitment formed early, before defensive behavior became self-reinforcing—implying that early intervention and coordination are essential when deploying LLM agents in financial systems.

Financial AI safety requires new evaluation frameworks: The paper demonstrates that current approaches to AI safety may be inadequate for financial applications, calling for system-level evaluation methods and interaction design as central priorities for responsible deployment of LLMs in financial decision-making.