Every serious proposal for governing AI agents ends in: liability, mandatory insurance, transparency requirements, monitoring. The MIT FutureTech Delphi study of 272 experts lands there too. It's the right list. It also assumes something that doesn't exist. Verification.
An underwriter can't price a risk it can't verify. A regulator can't enforce a transparency rule against a system that reports on itself. A liability claim needs a record of who did what, independent of the party being sued. Each instrument runs on the same missing input: a tamper-evident account of what an AI agent did, bound to a responsible actor.
That verification is a keystone. Without it, the governance stack is a set of rules with nothing underneath to enforce against.
The study measured the demand. Experts put 18 of 24 risks at more than a 10% chance of catastrophic harm, and judged the mechanisms to address them "nascent or absent." Supply is the open problem. Proof of execution, verifiable records of an agent's decision chain, is what fills it.
Full disclosure, I was one of the contributors in this study.
June 4, 2026
