A new paper from researchers including Eric Osterweil at George Mason University makes a narrow, testable case for using DNS as the discovery layer for AI agents, and backs it with data rather than just architecture diagrams. The core problem: as agents move from closed enterprise deployments to the open internet, they need a way to find each other, verify who they're talking to, and do it fast, across organizational boundaries that don't share a common registry. The paper argues three criteria will decide which discovery approach wins: whether the metadata format is complete enough to carry trust signals, how complex the lookup process is, and how it performs under real network conditions. To test the idea, the authors analyzed 119,757 real-world service endpoints and 62,739 indexed MCP server entries, then measured whether the metadata an agent needs to discover, authenticate, and connect to another agent could be packed into a single DNS resource record. It could — and it fits inside one DNS UDP response, meaning a lookup can resolve in a single round trip rather than the multiple back-and-forth calls that registry-based or crawling-based discovery schemes typically require. That single-RTT result is the paper's most concrete contribution. Most agent discovery proposals so far — capability registries, well-known URI files, centralized directories — assume some multi-step negotiation before an agent can even confirm another agent exists and is trustworthy. Reusing DNS instead means leaning on infrastructure that already resolves billions of queries a day, has established caching and propagation behavior, and doesn't require agents or platforms to adopt a new lookup protocol from scratch. The paper is explicit that this is a first cut at the problem, not a finished standard: it's making the case that DNS is *suited* to the job by testing the metadata-completeness question empirically, while lookup complexity and transaction performance across production traffic remain to be validated further. It doesn't resolve who runs the authoritative zones for agent records, how disputes over agent identity in DNS get adjudicated, or how this interacts with the handful of competing discovery proposals already circulating (capability registries, IETF drafts, platform-specific directories). What it does show is that the technical objection — that agent metadata is too rich to fit in a DNS record — doesn't hold up against real deployment data. If DNS-based discovery gains traction, the practical effect is fewer round trips before an agent can start a negotiation, which matters more as agent-to-agent calls scale into the billions and shave milliseconds start adding up.