The paper introduces FRAIL, a framework that places LLM agents in three dynamic financial environments: bank runs, debt rollover, and reward crowdfunding. Across seven leading LLMs, agents make individual decisions that collectively shape the financial state, mimicking real-world coordination problems.
In baseline simulations, 77% of bank run episodes and 83% of debt rollover episodes ended in failure, despite the environments being viable if participants cooperated. Crowdfunding was more forgiving, with 23% failing to launch. These failures occurred across all model families, not from a single model.
The authors tested three interaction mechanisms—compensated commitments, centralized commitment agreements, and participant-led coalitions—all of which improved aggregate outcomes, though no single mechanism worked best across all settings. Successful stabilization depended on broad commitment forming early, before defensive behavior became self-reinforcing. This shows that individually capable agents do not automatically create safe financial systems, highlighting the need for system-level evaluation and interaction design.