A Registration Framework for AI Agents

Kevin Werbach at Wharton has a paper, forthcoming in the Villanova Law Review, that treats the AI agent accountability problem as a registration problem. Agents that sign contracts, move money, and talk to third parties are hard to hold responsible, because the legal system has no reliable way to identify them.

His answer borrows from corporate registration law and runs in three tiers. Basic registration assigns an agent a unique identifier, so it becomes visible to the legal system at all. Standard registration turns the agent into a legal object with defined attributes and grants protections that make signing up worth it. Full registration asks for a surety bond posted through a smart contract and a Ricardian contract for resolving disputes. Werbach argues the tiered protections draw participation without a government mandate, and the records they leave behind give an AI insurance market something to price against.

The part that is interesting to me is the shift from philosophy to plumbing. Identity comes first. Before anyone can argue about whether an agent should be allowed to do a thing, there has to be a durable way to name it and attach a record to it. Capability without a registered identity is action nobody can trace. The fight is over who builds the registry and whose attributes it decides to trust.