Review · November 16, 2026
DNS for AI Discovery (DNS-AID) with SVCB Records
The Problem
An agent wants to call another agent. It needs the endpoint, the agent protocol (MCP, A2A), and a pointer to the agent's descriptor. Today that comes from configuration or a third-party directory.
DNS-AID publishes it in DNS, in records the agent's own organization controls. It uses SVCB records and defines no new record types.
How It Fits In
This is a candidate for the DAWN charter's Protocol deliverable. The charter says the working group "will consider DNS and mDNS as mechanisms upon which to build a discovery protocol", so the draft is on target there.
It is a -00 under this name, but not new work. Other DAWN drafts still cite it as draft-mozleywilliams-dnsop-dnsaid-02.
The draft covers two of its three use cases (Section 3):
- Known agent. One SVCB query at the agent's name. In scope.
- Known organization. One SVCB query at
_index._agents.<domain>, which points to a registry. In scope. - Known capability. Search across organizations. Out of scope.
This split lines up with the charter. An AI resource may be "a service such as a catalog that will further the discovery process", so the organization index is in scope. The charter excludes "open-ended search", which is the third case.
Two charter lines are a harder fit:
- Work items "will provide integrity of all the discovery information exchanged". DNS-AID makes DNSSEC a SHOULD and says it is "deployable without DNSSEC" (Sections 1.1 and 6.4).
- The Protocol deliverable includes "the definition of the minimum public discovery information". This draft defines none of its parameters itself.
So how does it differ from the other candidates?
- DAN defines two new record types, AIDISCA and AIINDEX. Certificate data sits inside the agent record, the index lists agent names in DNS, and DNSSEC is a MUST. DNS-AID reuses SVCB and TLSA, and its index is only a pointer to something outside DNS.
- ARD (architecture) is HTTP-first. The manifest at
/.well-known/ard.jsonis mandatory, and DNS is optional. ARD's initial contact draft already uses the_index._agentsSVCB record from DNS-AID as one way in. These two are layers, not rivals.
Summary
An agent lives at an ordinary name like agent-name.example.com. An SVCB record there carries alpn, port, address hints, and the agent parameters proto, proto-ver, path, well-known, desc-ref, and desc-ref-digest (Section 3.1, Table 1).
None of those agent parameters are defined in this draft. They come from Common Agent DNS Discovery Parameters and a DNSOP individual draft for well-known.
The organization index returns a host and a path. What is served there, in what format, over what protocol, is out of scope (Section 3.2).
TLS authentication is DANE if a TLSA record exists and WebPKI otherwise (Section 6.2).
Interoperability
Significant interoperability concerns
The parameters have no code points. Section 7.1 defers registration to the common parameters draft. That draft's IANA section says it "will be revised". Section 6.3 admits implementations may use different private-use numbers for the same key, so mandatory=proto "will not work reliably".
The TXT fallback is not a protocol yet. Section 4 says TXT at the same owner name carries the same keys. It does not say when a client queries TXT, or which record wins if both exist and disagree.
The TLSA owner name is defined two ways. Section 6.2 says _443._tcp.<owner>. Section 3.2 says to use the TargetName. These differ whenever the agent is hosted elsewhere, which is one of the draft's main examples.
Mechanisms that could be removed
Section 5 lists nine experimental items: EDNS(0) query hints, billing and jurisdiction parameters, connect-class, enroll-uri, a domain control challenge, and more. None are part of the "minimum public discovery information" the charter asks for. The domain control challenge and enroll-uri lean toward identity and trust management, which the charter excludes. Move them to an appendix or drop them.
Section 7.2 asks IANA for _agents-challenge on behalf of one of those experiments. That request should wait.
Writing Quality
The core is short and readable. The three use cases are a good frame.
The document has not been cleaned up since its rename:
- The header says "dnsop". The mailing list is
[email protected]and the repository isgithub.com/USER/REPO. - Section 5 says it replaces a section "in earlier revisions". This is -00.
- Section 5.9 discusses "the TXT record-based form of
_index._agents". No such form is defined.
Authors
From their Datatracker profiles:
- Jim Mozley (Infoblox). No RFCs. 4 active drafts, including both drafts this one depends on.
- Nic Williams (Infoblox). No RFCs. 3 active drafts, all in DAWN.
- Behcet Sarikaya (Unaffiliated). 12 RFCs, mostly mobility, IPv6, and low-power networks. 3 active drafts.
- Roland Schott (Deutsche Telekom). 3 RFCs, on RTCP reporting and ALTO. 7 active drafts.
- Jeffrey Damick (Amazon). No RFCs. 2 active drafts, this one and ARD discovery.
- Scott Courtney (GoDaddy). No RFCs. 2 active drafts, this one and Agent Name Service v2.
Questions for the Working Group
So what should Thursday's session decide?
- New record types or SVCB, which is the real DAN versus DNS-AID choice?
- Does DNSSEC as a SHOULD meet the charter's integrity requirement, when DAN says MUST?
- Should DNS-AID and the common parameters draft be one document, since neither is implementable alone?
- Should DNS-AID and ARD be reviewed as one candidate, if the index points to an ARD manifest in practice?