Update review · November 16, 2026
DNS-Based Agent Naming (DAN) -01, the Index Moves to _agents
What changed in -01?
Very little. The document is still 21 pages. The diff is one design change, one new appendix section, and spelling fixes.
The -00 review covered two new DNS record types, AIDISCA for one agent's metadata and AIINDEX for a list of agent names. Both wire formats are unchanged. So are the IANA requests and the security considerations.
The one design change
AIINDEX moved. In -00 it lived at the zone apex, example.com. In -01 it SHOULD be published at _agents.example.com (Section 3). The example in Section 3.3 and the client lookup in Section 4.1 were updated to match.
This is a good change. The index now sits next to the AIDISCA records it lists, under one label.
It also brings DAN closer to its main competitor. DNS-AID publishes its organization index at _index._agents.<domain>. The two drafts now agree on the _agents label and disagree on almost everything under it. DNS-AID uses an SVCB record that points to an external registry. DAN uses a new record type that lists agent names inline.
The edit is incomplete:
- Section 1 still says AIINDEX advertises "a DNS zone's available AI agents." Section 1.2 still says discovery is "on a per-zone basis." A record at
_agents.example.comis per name, not per zone. - Section 4.1 tells the client to query "the
_agentsname of the specific service domain." That term is not defined. - Section 3 still allows other locations with a MAY. Section 4.1 only tells the client about
_agents. An index published anywhere else will not be found.
What else is new?
Section A.1 is new. An orchestration or hosting system may provision the records on behalf of the agent. Section A.2 gained a sentence on coordinating record updates during deployment, migration and removal.
Nothing was removed.
Was the previous feedback addressed?
Partly addressed:
- Naming convention. The index now has one recommended location that a client can construct from a domain name. The AIDISCA convention is still a SHOULD with "other naming conventions MAY be used" (Section 2.3).
Not addressed:
- Free-form capabilities. Section 2.2.4 is unchanged. Capability strings are still agent-defined, with no vocabulary, no registry and no matching rules. A client looking for
hotel-bookingstill misseshotel_booking. - Extensions mechanism. Still there, in both records, with one defined code (Agent Card).
- AIINDEX and enumeration. Section 7.5 is unchanged. There is still no analysis of what a signed, public list of agent names exposes.
- The arXiv paper. Sections 1.4 and 6 still rest their size and atomicity arguments on arXiv:2606.02314.
- Endpoint verification scope. Section 7.3 changed one word, a spelling fix. The endpoint can still be in any domain, with no guidance on what a client should conclude from that.
- Multiple AIDISCA records. Section A.4 still has no rule for how a client picks between an MCP record and an A2A record at the same name.
Section 5 still opens with a leftover note: "Goal: clients want to resolve and connect to AI agents..."
Background of the authors
The author list did not change. It is the same four Verisign authors in the same order. See the -00 review for their background.
Questions worth raising
DAN has a 10 minute slot on Thursday. The charter says solutions will be built "using existing IETF protocols" where possible, and that the working group "will consider DNS and mDNS" as the base.
- New record types or existing ones? DNS-AID does the same job with SVCB and TLSA, and Section 5 argues for atomicity without naming DNS-AID or showing a case where SVCB fails.
- Should the index live in DNS at all? The charter allows "a service such as a catalog that will further the discovery process," so the choice is between DAN's inline list, DNS-AID's pointer to a registry, and the ARD architecture manifest at
/.well-known/ard.json. - Can DAN and DNS-AID agree on the index name?
_agentsversus_index._agentsis a small gap to close now, whichever record type wins. - Is a free-text capability list part of "the minimum public discovery information"? The charter puts catalogs of capabilities and semantic matchmaking out of scope, so a string no client can match on needs rules or needs to go.