The Scope Creep Automated reviews of new AI-related working group submissions

Review · August 21, 2026

DNS-Based Agent Naming (DAN): AIDISCA and AIINDEX Resource Records

draft-seethiraju-dawn-dan -00View on Datatracker →

Problem Statement

AI agents operating across different organizations and internet domains have no standardized mechanism to discover one another's network endpoints and capabilities. This draft addresses the bootstrapping phase only — getting an agent's address and a coarse capability description before any connection is made — leaving richer runtime semantics to the protocols themselves (MCP, A2A). It proposes using DNS as the publication layer via two new resource record types.

How the Spec Fits In

This is a DNS infrastructure draft, not an application-layer protocol spec. It sits in the DAWN space (AI agent discovery) and draws heavily from the DNS-based Authentication of Named Entities (DANE) architecture defined in RFC 6698. The approach is essentially: apply DANE's DNS-anchored certificate binding pattern to AI agent metadata. The spec:

  • Introduces two entirely new DNS Resource Record types, requiring IANA allocation and DNS implementation changes
  • Reuses RFC 6698 certificate association fields verbatim, without modification
  • Depends on DNSSEC (RFC 4033–4035) as a hard requirement, on DANE/TLSA semantics (RFC 6698), and on DNS-over-TCP fallback (RFC 7766)
  • Does not depend on any OAuth or application-layer identity RFC — it operates one layer below

This positions it as a foundational building block for agent-to-agent connectivity rather than an end-to-end solution.

High-Level Summary

The draft defines two new DNS record types. AIDISCA (AI Discovery and Connection Association) is published at an agent's DNS name — such as booking._agents.example.com — and carries four pieces of information in a single atomic record: a protocol identifier (e.g., MCP or A2A), a free-text capability string, the agent's service endpoint URI, and certificate association data modeled exactly on TLSA records. An organization running an agent publishes an AIDISCA record; another agent queries for it, DNSSEC-validates the result, then uses the endpoint and certificate data to establish an authenticated connection.

AIINDEX is published at a zone apex (e.g., example.com) and acts as a directory — it lists the DNS names where AIDISCA records in that zone are published. This lets a consumer enumerate all agents an organization has chosen to advertise without DNS crawling.

The client processing model has three phases: optional discovery via AIINDEX (skipped if the agent's DNS name is already known), metadata retrieval via AIDISCA, and endpoint connection with certificate verification per RFC 6698.

The spec explicitly scopes out semantic search, capability matching, ranking, and agent selection — those are considered external concerns. DAN is strictly the publish-and-lookup layer.

Interoperability Assessment

Significant Interoperability Concern: Free-Form Capabilities

The Capabilities field is defined as a comma-separated list of UTF-8 strings with no defined vocabulary, no registry, and no reference to any controlled namespace. The spec justifies this by noting that AI agents use free text. This directly undermines interoperability: two implementations publishing and consuming capability strings will have no shared basis for interpretation. If a client queries for agents with capability "hotel-booking" and the operator published "hotel_booking" or "Book Hotels," discovery fails silently. A minimal controlled vocabulary, a URI-based naming convention, or a Specification Required registry would address this without constraining agents from adding richer semantics later.

Naming Convention Underspecified

The _agents label convention for where AIDISCA records are published is a SHOULD, not a MUST, and other conventions are explicitly permitted. This means clients querying an unknown domain cannot reliably know where to look for AIDISCA records without first consulting AIINDEX — which itself requires knowing the zone apex. In practice, competing naming conventions will emerge across implementations.

Mechanisms That Could Be Removed or Tightened

The Extensions mechanism (EDNS(0) CLV encoding within DNS record data) adds considerable implementation complexity for what is currently a single defined extension (Agent Card pointer). The "Agent Card" extension could be deferred to a future document once the base record is deployed and actual needs are understood — adding an extension mechanism before there's demonstrated need for multiple extensions is premature.

The AIINDEX record's value over zone enumeration is only meaningful if DNSSEC is already deployed (which it mandates), but its security properties against zone enumeration are not analyzed. It may actually increase discoverability beyond what zone operators intend when combined with denial-of-existence records.

Writing Quality

The writing is reasonably clear for a -00 individual draft. Wire format diagrams are correctly formatted. RFC 6698 fields are reused by reference without re-specifying their semantics, which is appropriate. However, there are several gaps:

  • Section 1.4's argument for atomic vs. separated metadata is asserted but the empirical backing ("companion analysis paper arXiv:2606.02314") is an arXiv preprint, not an IETF document. IETF specs should not normatively rely on arXiv references.
  • The endpoint verification scope section (§7) notes that the service endpoint need not share any domain relationship with the AIDISCA record, but does not specify how implementations should reason about this trust relationship — it is security-relevant and under-specified.
  • "SHOULD use _agents label" without a corresponding MUST-level fallback creates ambiguity that the spec does not resolve.
  • The spec does not define how multiple AIDISCA records at the same name for different protocols are selected or prioritized by a client.

The document is appropriately concise for its scope.

Author Background

Ramachandra Rao Seethiraju (Verisign) — No published RFCs. One active draft (this document). New to the IETF.

Sameer Thakar (Verisign) — No published RFCs. One active draft (this document). New to the IETF.

Karthik Shyamsunder (Verisign) — No published RFCs. One active draft (this document). New to the IETF.

Eric Osterweil (Verisign) — Three published RFCs: RFC 7094 (IP Anycast architecture, 2014), RFC 7682 (Internet Routing Registries, 2015), RFC 7908 (BGP Route Leak classification, 2016). Experience spans GROW, SIDR, and DNSOP working groups. The most experienced IETF participant among the authors, though his published work is in routing and DNS operations, not AI or application-layer discovery.

This is a Verisign-internal team effort with one experienced DNS/routing contributor and three IETF newcomers. Verisign's institutional DNS expertise is relevant and credible for this type of work, but the team has no prior experience shepherding a draft through the RFC process.