Achieving Agent Accountability with Verifiable Agent Insurance (VAI)

The researchers propose the Verifiable Agent Insurance (VAI) framework, in which agent providers financially guarantee a set of programmatically specified policy commitments; the commitments, together with the escrowed funds, are written into smart contracts on an immutable ledger such as a blockchain to prevent tampering. If a user discovers a violation, they can trigger a verifier to review the execution trace, and after verification the provider's escrowed funds are paid to the reporting user. Traces are recorded through the shared tool execution paths of the agent harness, without instrumenting each tool individually. The team completed an end-to-end implementation on an Ethereum layer-2 network and evaluated coverage on three agent benchmarks, two of which were relabeled with task-level policy compliance labels. Because many current agents do not output user authorization and safety-related information as explicit structured data, VAI detected 42.2% of violations among runs containing violations on the three benchmarks; after augmenting agents to explicitly output such objects, violations were detected in 83.0% of traces. The researchers accordingly recommend co-designing agents and safety policies, explicitly outputting structured records of user authorization and safety-related operations to improve auditability.