Most agent identity proposals — including the ones this feed has already covered from CrowdStrike, the AWS/Zscaler/Ping/Defakto auth spec, and various trust protocols — assume an agent presents a credential and a verifier reads it. A May 2025 paper from Ken Huang, Vineeth Sai Narajala, and coauthors takes a different tack: it argues agents should be able to prove they satisfy a policy without exposing the underlying credential at all. The paper's core complaint is that OAuth, OIDC, and SAML were built for humans logging into one service at a time, not for autonomous agents chaining through dozens of tool calls with ephemeral, delegated scopes. Their proposed framework builds agent identity on decentralized identifiers and verifiable credentials that encapsulate an agent's capabilities, provenance, behavioral scope, and security posture. Layered on top is an Agent Naming Service for discovery, fine-grained access control, and a global session layer for consistent revocation across protocols that don't otherwise talk to each other. The part worth flagging is the zero-knowledge proof layer. Instead of an agent handing a downstream service its full credential — organization, clearance level, training data provenance, whatever — it would generate a proof that it meets a specific policy condition ("cleared for financial data access," say) without revealing anything else in the credential. The paper frames this explicitly as enabling privacy-preserving attribute disclosure and verifiable policy compliance, which matters once agents are routinely presenting credentials to services run by parties with no reason to see the rest of the agent's profile. This is a meaningful gap in the current wave of agent-auth proposals. Most treat "verify the credential" and "protect the credential's contents" as the same problem solved the same way — check the signature, read the claims. A ZKP layer separates them: the verifier gets a yes/no on a specific claim, not a dossier. For multi-agent systems where an agent might pass through several organizational boundaries in a single task, that's the difference between minimal disclosure and every hop accumulating a copy of the agent's full credential history. The paper is explicit that this is architecture, not deployment — it lays out the operational lifecycle and design tradeoffs rather than a working implementation. But as a design target for the identity layer, selective disclosure via ZKPs is a more precise answer to "how much should a verifier actually get to see" than the credential-presentation model most current proposals default to.