Most proposals for agent identity converge on the same answer: make every agent traceable to a real person, full stop. A new paper documents a different approach already built as national infrastructure in China, scheduled for public launch in Q3 2026. The researchers describe how the emerging infrastructure for AI-agent identity has converged on making every agent identifiable, in industry practice and research proposals alike. The Chinese system takes a different path: an agent is tied to a verified legal principal — a real person or company responsible for it — without that principal's identity being exposed to any business, platform, or counterparty the agent interacts with. The mechanism, which the authors call split-knowledge binding, keeps re-identification possible only through a legal process that requires compelling two separate government agencies at once, neither of which holds enough information alone to unmask the person behind an agent. So a merchant, a platform, or another agent negotiating with yours sees a verified-but-anonymous identity; only a court order routed through both agencies can pierce that. The authors are upfront about the limits of this design. The separation is structural and procedural, not cryptographic, and a state powerful enough to compel both agencies simultaneously can still re-identify anyone. That's a real caveat — this isn't zero-knowledge cryptography preventing re-identification even from the state itself, it's a bureaucratic separation-of-powers argument, closer to how wiretap warrants or dual-key nuclear launch systems work than to a cryptographic privacy guarantee. What makes this notable against the rest of the agent-identity landscape is the framing choice. Nearly every Western proposal — biometric binding, on-chain identity, verifiable credentials — treats accountability and identifiability as effectively the same problem: you can only hold an agent's owner accountable if everyone downstream can see who that owner is. China's system, at least as documented here, tries to decouple those two things: accountability exists as a legal backstop, but day-to-day visibility doesn't leak to every counterparty an agent talks to. Whether this holds up depends entirely on how independent those two government agencies actually are in practice, something the paper itself doesn't resolve and which outside observers won't be able to verify until the system launches. But as a design point in the space of agent-identity architectures, it's a genuinely different tradeoff than the industry consensus of "just make agents identifiable to everyone."