Review · November 16, 2026
Common DNS Discovery Parameters for Agents
The Problem
DAWN has two DNS-based discovery proposals. DNS-AID publishes agents in unicast DNS with SVCB records. Zero-Configuration Agent Discovery advertises them on the local link with mDNS and DNS-SD TXT records. Each had its own key names.
This draft defines one set of nine keys for both.
How It Fits In
This is an individual draft at -00. Its four authors are the lead authors of the two drafts it unifies, two from each.
Do the two drafts actually use it? Yes, and both cite it normatively.
- The mDNS draft (-01) dropped its own keys and uses all nine. It has changed a lot since our review of -00.
- DNS-AID uses seven of the nine in its SVCB table. It never mentions
titleortxtvers.
Both drafts also hand their IANA work to this one. DNS-AID says its SVCB parameters "will be registered via" this draft. The mDNS draft says a registry for proto values "is a matter for" this draft.
Which charter deliverable is this? Protocol. That deliverable includes "the definition of the minimum public discovery information", and this draft is a candidate for exactly that part. It is not a protocol on its own.
The fit with the charter is good. The charter says the working group "will consider DNS and mDNS". Leaving capabilities out matches the charter, which puts "catalogs of capabilities" out of scope.
It depends normatively on an unadopted individual draft, draft-mk-dnsop-svcb-well-known, for the well-known key.
Summary
The draft says a DNS record should answer three questions: how do I connect, what am I connecting to, and is what I cached still current (Section 1). Everything else comes from the agent's descriptor after connecting.
alpn: HTTP versions only, such ash2,h3. Never the agent protocol.txtvers: TXT format version. Must be the first key. TXT only.proto: one agent protocol per record, such asa2aormcp.proto-ver: a version hint in the protocol's own syntax.desc-ref: an opaque reference to the agent's descriptor, such as a URN.desc-ref-digest: a hash used to tell whether a cached descriptor is stale.path: the HTTP path of the endpoint, such as/mcp.well-known: the descriptor's suffix under/.well-known/.title: an unverified display name.
An agent that speaks two protocols is advertised twice, once per protocol. A client links the two records by matching desc-ref values (Section 3.4).
Interoperability
Significant interoperability concerns
The SVCB form cannot interoperate yet. Section 3.4 says SVCB uses private-use numeric codes "until IANA assigns". The draft does not list which numbers. Section 6 is a placeholder. The draft says TXT is the fallback, but DNS-AID tells requestors they MUST try SVCB first.
It is unclear what desc-ref-digest hashes. Section 3.2.5 calls it a "digest of the descriptor reference". That would be a hash of a string already in the record. The mDNS draft says it is a digest of the AgentCard. Nothing says which bytes are hashed or how a JSON document is canonicalized.
desc-ref is an unconstrained, unauthenticated string. Any octets except NUL (Section 3.2.4). Comparing it for equality is what links an agent's A2A and MCP records. Section 5 covers spoofed titles and nothing else. The charter says work items "will provide integrity of all the discovery information exchanged".
Smaller gaps:
- There is no
protoregistry. Section 3.2.3 says values need not be registered anywhere. txtversis a MUST in TXT and must come first (Section 3.2.2). DNS-AID's TXT section never mentions it.- Section 3.2.8 says
well-knownis one suffix. Appendix B.14 and DNS-AID's table allow several.
Mechanisms that could be removed
desc-ref-digest, until the draft says what it hashes.titlein SVCB. DNS-AID does not use it.
Writing Quality
This reads like a merge still in progress. Appendix B is the best part. It records fourteen rejected alternatives with reasons.
The body needs an editing pass:
- Section 4, Open questions, is empty.
- Section 3.2.5 says
_agent._tcp.local.. Everywhere else it is_agents._tcp. - Appendix B.12 says
well-knownis the one key longer than nine characters.desc-ref-digestis fifteen.
Authors
From their Datatracker profiles:
- Frank Brockners (Cisco). 10 RFCs, mostly In Situ OAM. 7 active drafts. Listed as a DAWN chair.
- Loránd Jakab (Cisco). 1 RFC, RFC 7215 on LISP deployment. 3 active drafts.
- Nic Williams (Infoblox). No RFCs. 3 active drafts.
- Jim Mozley (Infoblox). No RFCs. 4 active drafts. Listed as chair of TDD.
Questions for the Working Group
So what should the session decide when it reviews proposals against the deliverables?
- Is this key set the charter's "minimum public discovery information", given it has no key for the type or services the charter lists as examples?
- Where does the integrity the charter requires come from, especially on mDNS?
- What does
desc-ref-digesthash, exactly? - Who registers the SvcParamKeys and the
protovalues, and when?