Left to their own devices, AI agents panic just like people do.
Researchers tested seven leading LLMs in FRAIL, a framework that simulates three financial scenarios: bank runs, debt rollover, and reward crowdfunding. In each, one agent's decision reshapes the conditions everyone else faces. No agent was instructed to cause a collapse. Even so, 77% of baseline bank-run episodes and 83% of debt-rollover episodes ended in failure.
The researchers then tried three fixes: compensated commitments, centralized commitment agreements, and participant-led coalitions. All three improved outcomes, but none won across every scenario. The mechanisms that worked shared one trait: broad commitment formed early, before defensive behavior became self-reinforcing.
Rational agents, individually correct, still produce collective failure. Same story as 2008, new cast.