SAGE-Fin: A New Framework for Runtime Governance of Financial AI Agents

A recent research article presents SAGE-Fin, a contract tailored for finance that facilitates the transfer of authority, ensuring that the proposed impacts of AI agents are controlled during runtime, not just their textual outputs. This framework organizes proposals into typed candidates bound to adapters and monitors unfulfilled institutional obligations as coverage debt. Authority is contracted based on the current market, account, policy, and dialogue context, necessitating a precise artifact receipt that corresponds to the type of the consuming response, execution, or policy adapter. The system guarantees that evidence and workflow advancements cannot replace effect authority, with prior approvals being re-evaluated following state changes. The authors evaluated SAGE-Fin using a catalog of 616 cases, producing 3,080 outputs from five deterministic specifications. A label-isolated harness achieved perfect parity at 616/616 for binary reference-prototype, including 3/3 named response-gate fixtures. The paper can be accessed on arXiv with the identifier 2608.09025.