A new paper sketches out something like a DNS for AI agents - infrastructure for finding and vetting them before they start taking instructions from each other.
The proposal, posted to arXiv, splits the job into three layers. An Agent Root governs which registries can be trusted, an Agent Registry handles agent sign-ups and metadata, and an Agent Resolver does the actual lookup work, matching a request for a capability to an agent that can provide it. Identity runs through a registry-suffix-anchored composite scheme, which ties an agent's own identifier to a trusted registry suffix - conceptually similar to how a domain name anchors identity on the web. A dual-certificate, multi-level authentication layer sits on top to verify agents are who they claim to be. In a prototype, the authors report 58-millisecond average registration times, 25-millisecond discovery times, and throughput above 19,000 registrations and 29,000 discovery requests per second.
The gap this targets is real. The paper itself notes that current agent protocols mostly cover tool invocation and agent-to-agent messaging, not how an agent finds a trustworthy counterpart before that conversation even starts. As autonomous agents increasingly transact without a human in the loop, that missing discovery-and-trust layer starts to look like the early internet's need for DNS and public-key infrastructure - minus decades of adversarial hardening.
Worth remembering: this is one team's preprint with self-reported benchmarks, not a deployed standard. DNS and PKI took years of real-world abuse to get right. An Internet-of-Agents trust layer will need the same scrutiny before anyone should lean on it to vet who, or what, they are actually talking to.