Published February 2026 9 min read

Why DNS Fails for AI Agents

DNS maps names to IP addresses. Agents need capability discovery, trust verification, and protocol negotiation — none of which DNS provides.

Technical Infrastructure

DNS: A 1983 Masterpiece

The Domain Name System is one of computing's great achievements. Designed by Paul Mockapetris in 1983, DNS provides a globally distributed, hierarchical namespace that maps human-readable domain names to machine-readable IP addresses. It's fast, resilient, and handles billions of queries daily.

But DNS was designed for a specific purpose: locating servers that host documents. When you type example.com, DNS returns an IP address. Your browser connects, requests a page, and the transaction ends. This model has served the web brilliantly for four decades.

The problem? AI agents aren't documents. They don't sit at fixed endpoints waiting for requests. They're dynamic, autonomous entities that discover peers, negotiate protocols, verify trust, and delegate tasks — all in real time.

Five Ways DNS Falls Short

1. Propagation Latency

DNS updates propagate through a hierarchy of caches and resolvers. Research shows end-user visibility can stretch to 24–48 hours in worst cases, with typical propagation taking minutes to hours. When an agent spawns a helper agent that needs to be discoverable immediately, minutes are an eternity.

As Professor Raskar noted during an AWS DNS outage: "DNS glitch at AWS today for websites. Imagine if we had the same for the Internet of AI Agents."

2. No Capability Metadata

A DNS record tells you where something is — an IP address, maybe a port. It tells you nothing about what it can do. When your travel-booking agent needs to find a flight-search specialist that supports A2A protocol, speaks English and Japanese, and has a verified performance SLA, DNS has no mechanism to express or query any of this.

3. Trust Is Domain-Level Only

TLS certificates prove domain ownership — that example.com is controlled by Example Inc. But agent trust requires far more: behavioral history, capability attestations, compliance certifications, and cryptographic proof of code integrity. As the NANDA Beyond DNS paper argues, trust must be agent-level, not domain-level.

4. Privacy Exposure

Every DNS query reveals what you're looking for to network observers. When a pharmaceutical company's research agent queries for a genomics specialist, that search pattern itself leaks strategic intent. NANDA's dual-path resolution architecture addresses this with anonymous lookup paths through neutral third-party hosts.

5. Static Addressing in a Dynamic World

Agents migrate between runtimes, auto-scale across regions, and update capabilities continuously. DNS was designed for relatively static mappings — a domain name pointing to one or a few IP addresses. Agents need dynamic, adaptive routing that reflects their real-time state.

The Scale Problem

These limitations compound at scale. Consider the numbers: DNS currently handles roughly 400 million domain registrations globally. The Internet of AI Agents envisions billions to trillions of agents — each registering, updating capabilities, rotating endpoints, and revoking trust in real time.

~400M
DNS Domains Today
Trillions
Projected AI Agents
Hours
DNS Propagation
< 1s
Agent Discovery Target

The NANDA Index addresses this with an ultra-lean record format — AgentAddr records of ≤120 bytes — that separate static identity from dynamic metadata. This architectural decision reduces write overhead by approximately 10,000× compared to DNS while enabling sub-second global resolution.

What Replaces DNS for Agents?

The answer isn't to deprecate DNS — it continues to work well for its intended purpose. The answer is to build a complementary discovery layer purpose-built for agents, just as broadband infrastructure was built alongside (not instead of) the telephone network.

The NANDA protocol provides this layer through three interlocking systems:

  • The NANDA Index — a decentralized "Quilt" of registries providing global agent discovery without a single point of control
  • AgentFacts — rich, cryptographically signed metadata encoded as W3C Verifiable Credentials
  • The Adaptive Resolver — a dynamic resolution layer handling real-time endpoint discovery, federation, caching, and failover
The resolution flow. AgentName → NANDA Index → AgentAddr → AgentFacts → Agent Endpoint — analogous to DNS resolution but with built-in trust verification, capability matching, and privacy-preserving lookups at every step.

Further Reading

The architectural case for moving beyond DNS is laid out in detail across several NANDA research papers:

Continue Reading

Coming Soon

By Invitation Only