Agentic AI Systems and Financial Stability: From Model Risk to Systemic Risk

01 ABSTRACT

This work argues that when many decision-makers depend on the same infrastructure, distress is no longer approximately idiosyncratic and diversifiable, so the binding concern shifts from the model risk of a single deployment to the systemic risk of the population. The authors formalize this across six mathematical settings: a shared foundation model is a non-diversifiable common exposure whose expected-shortfall floor does not vanish as the fleet grows, and a percolation threshold governs contagion. The authors' central claim is that the systematic component of agentic risk cannot be diversified away, detected away, or reversed, and only ex-ante structural prevention moves it.

02 KEY FINDINGS

1. A shared foundation model is a non-diversifiable common exposure whose expected-shortfall floor persists as the fleet grows, with a percolation threshold governing contagion.
2. Under a lever-assignment axiom that review cannot intercept an irreversible action, no non-preventive control satisfies a coherent tail constraint.
3. Modeling runtime guardrails as partially observed stochastic control yields an observability trichotomy, a detection floor, and, under commit exogeneity, a runtime-impossibility corollary.