Header image source: Working Title | LinkedIn via LinkedIn via Google — cropped to 16:9 and colour-adjusted.
Key takeaways
- LLM agents cause 77% simulated bank runs without malicious intent
- Multi-agent financial systems show inherent fragility from interactions
- Future AI crises may unfold at machine speed beyond human intervention
FRAIL just dropped a number that should make anyone building financial AI sit bolt upright: 77%. That’s the fraction of simulated bank runs that collapsed when LLM agents were left to their own devices. No hackers, no rogue traders—just seven different language models making what looked like reasonable decisions. Debt rollover fared even worse: 83% failure rate. They’re the first hard evidence that multi-agent financial systems built on today’s LLMs are inherently fragile. And that fragility doesn’t come from the agents themselves—it emerges from their interactions.
The Experiment: FRAIL’s Financial Petri Dish
It’s a controlled experimental framework designed to isolate three classic financial coordination problems: bank runs, debt rollover, and reward crowdfunding. Each environment is stripped down to essentials, removing noise so researchers could watch how LLMs behave when their decisions depend on each other.
In the bank-run scenario, agents play depositors deciding whether to withdraw cash. In debt rollover, they’re lenders choosing whether to refinance a borrower’s debt. In reward crowdfunding, they decide whether to contribute to a project that only pays out if enough others chip in. No agent is told to break the system. They’re just following prompts, optimizing for their own objectives. Yet in 77% of bank runs and 83% of debt rollovers, the system collapses anyway.
The researchers tested seven leading LLMs—the paper doesn’t name specific models. The failures weren’t limited to one model. They cut across architectures, sizes, and training data. This isn’t a bug in a specific LLM. It’s a feature of multi-agent financial systems built on current AI.
The Collapse Numbers: 77% and 83% Are Not Random
77% of bank runs failed. 83% of debt rollovers. They’re the baseline. It ran standard configurations with no added stress—and the system still fell apart most of the time.
Real-world bank runs are relatively rare events. But in FRAIL’s simulations, runs happen in 77% of cases without any external shock.
They’re just making binary choices based on limited information—and 83% of the time, the system locks up anyway.
Why This Happens: The Mechanics of LLM-Driven Fragility
The root cause isn’t malicious agents or bad code. It’s interdependence.
It just follows its prompt.
The study builds on prior work showing that interacting LLM agents can collude in simulated markets. It shows that even without collusion, even without adversarial intent, multi-agent financial systems can collapse under their own weight. This isn’t about bad actors. It’s about emergent fragility—the kind that arises when individual rationality leads to collective irrationality.
From Individual Agents to Systemic Risk: The AI Safety Blind Spot
They’re multiplicative.
But FRAIL shows that even well-aligned agents can destabilize a system just by pursuing their own objectives. The problem isn’t the agents. It’s the system.
The study’s innovation is extending failure-mode research to a unified framework for financial fragility. It’s not just about bank runs or debt rollovers in isolation. It’s about the underlying coordination problems that cut across all financial systems. And it’s not just about characterizing failures. It’s about testing solutions.
Stabilizing Mechanisms: What Works and What Doesn’t
The Bigger Picture: What This Means for AI in Finance
They’ll be systemic.
And if a collapse does happen, who’s responsible—the developers, the institutions, or the regulators who approved the system?
If AI-driven financial systems are prone to collapse, who bears the cost? In the 2008 crisis, it was taxpayers. In a future LLM-driven crisis, it might be depositors, investors, or entire economies. And because these systems operate at scale, the damage could be global.
Beyond Finance: Lessons for Multi-Agent AI Systems
FRAIL’s implications extend far beyond finance. Any domain where multiple AI agents interact—supply chains, social media moderation, autonomous vehicles—could face similar coordination failures. Imagine a fleet of self-driving trucks, each optimizing its own route, causing a traffic jam that no single agent intended. Or a group of content-moderation LLMs, each flagging posts based on its own rules, leading to unintended censorship.
The study is a wake-up call for anyone deploying multi-agent AI systems. It shows that safety isn’t just about individual agents. And right now, we’re building systems without understanding how the pieces interact.
Future research needs to go deeper. FRAIL’s environments are stylized—simplified to isolate coordination failures. But real-world financial systems are messy. They involve high-frequency trading, decentralized finance, and complex regulatory environments. Can FRAIL-like frameworks model these? And can they test interventions that go beyond classic financial stability tools—things like real-time oversight, dynamic regulation, or even AI-driven circuit breakers?
The Path Forward: Can We Fix This?
FRAIL proves that LLM-driven financial systems are fragile. The question is: can we make them resilient? The study’s interventions help, but they don’t solve the problem entirely. And each comes with trade-offs.
The bigger challenge is that fragility might be inherent to multi-agent systems. When agents interact, their decisions become interdependent in ways that are hard to predict. Even with perfect information, even with well-aligned objectives, the system can still collapse. This isn’t just an AI problem. It’s a systems problem.
The financial crises of the future may not be caused by human error or malice. They might be caused by AI agents doing exactly what they were trained to do—optimizing for their own objectives, unaware of the system-wide consequences. And because these agents operate at machine speed, the collapses could happen faster than we can blink.
We’re starting to see the problem. FRAIL is the first step toward understanding it. The next step is designing systems that don’t just survive but thrive—systems where coordination failures are the exception, not the rule. That will require new tools, new regulations, and a new mindset. Because right now, we’re building financial systems that are one bad decision away from collapse. And with LLMs in the driver’s seat, those decisions are happening faster than we can count. What happens when the next crisis unfolds at machine speed—and no human is fast enough to stop it?