Review · November 16, 2026
Registry-Assisted Resolution for Agents Without a DNS Anchor
The Problem
Most DAWN proposals assume the publisher controls a DNS zone. Alice at [email protected] does not. Neither does a workload with a SPIFFE ID or a DID.
This draft covers that case. The client already knows the subject's identifier and asks a configured registry where the subject's agent descriptor lives.
How It Fits In
This is an individual draft with a 5 minute slot in the protocol section of the DAWN session. It is a big retreat from -01. Section 1.6 says NandaIndex is now only a reference implementation. Search, ranking and registry-to-registry routing are all excluded.
So is a registry in scope for DAWN? Yes. The charter says the WG "will consider DNS and mDNS", but it does not require them. It says discovery "may use referrals, direct transfer, delegation, or indirection", and that the resource found may be "a service such as a catalog that will further the discovery process". An exact-lookup registry fits that. It is not the "indexing and searching across the wider Internet" the charter excludes.
Two charter exclusions sit close to the design: "identity management" and "trust evaluation methods beyond secure content exchange". The draft leaves enrollment out (Section 9.1). Its security still depends on checking proofs of control over email accounts, DIDs and workload identities. The charter also requires "integrity of all the discovery information exchanged". Here that integrity rests on things the draft does not define.
Which deliverable does it address? The Protocol one, through the HTTPS binding in Section 6. The draft calls that binding "experimental" and is unsure where it belongs. Appendix C asks whether it goes in the architecture document, the use cases document, or a separate protocol spec.
How does it compare to the other candidates?
- DNS-AID covers "known agent" and "known organization", both starting from a domain. Its Section 5.5 calls a consolidated registry "upper-layer work". This draft fills that slot for the known-agent case.
- DAN argues against this design. Section 5 of DAN -01 lists the costs of an external registry: extra latency, a separate reliability profile, and partial failure states.
- ARD already has registries, and it shares one principle with this draft: the registry host "establishes nothing about the publishers of the entries it holds". ARD registries do search. This draft does exact lookup only. It could be a lookup profile of ARD instead of a separate protocol.
Summary
The client sends POST /resolve with one identity and an identityType. The registry returns zero or more targets. Each target has a URL for an Agent Card or similar descriptor, a sequence number, issue and expiry times, and an authority object (Sections 5.2, 6.2).
Two rules shape the design:
- Provenance is not authority. The registry proves which registry answered. Separate Authority Evidence must show the subject approved the target (Sections 5.4, 10.7).
- No global root. Clients pick registries by local configuration. Registries never forward to each other (Sections 3.4, 6.6).
Interoperability
Significant interoperability concerns
Authority Evidence is not specified. It is the whole security argument, and it is an untyped bag. type is a free string with no registry. The examples use domain-control and account-provider-assertion with an evidenceUrl and no defined format behind it. The resolver MUST validate it (Section 7, step 8). Nothing says how. Two resolvers will not accept the same bindings.
The draft's own example breaks its main rule. In Section 8.2 the authority.issuer is https://registry.example, the registry itself. That is provenance presented as authority.
The same example undercuts the use case. Moon Bakery proves control with a file at https://moon-bakery.example/.well-known/example-proof. Section 2 defines a DNS Discovery Anchor to include "a well-known discovery resource under an authoritative domain". If the bakery can host that file, it has an anchor. DNS-AID's index record can also point at a hosted provider. The real gap is the email, DID and workload cases.
Nobody knows which registry to ask. There is no default registry, and the draft chooses not to define registry discovery (Sections 1.4, 3.3). A client that wants Alice's agent must already be configured with the registry Alice enrolled in. That fits the charter's "cooperating parties with trust relationships" context. Beyond it, everyone gets pushed toward a few large registries.
Registry provenance is undefined too. Section 5.6 allows a "response signature or authenticated-transport identity". No signature format is defined. In practice provenance means TLS.
Mechanisms that could be removed
- The
gatewaytarget class. Its behavior is not defined. maxResults,scopeandfallback. One subject, one current binding is enough for a first version.- Domain-name subjects. The DNS drafts cover them.
Writing Quality
The scoping is clear. The normative language needs work. Section 5.4 says the object "SHOULD contain" members that are then marked REQUIRED. The same boundaries are restated in the abstract, Sections 1.2, 1.4, 6.6, 15 and Appendix B. It could be half the length.
Authors
From their Datatracker profiles:
- Ramesh Raskar (MIT Media Lab). No RFCs. 5 active drafts.
- Samuel Sharaf (ProjectNanda). No RFCs. 2 active drafts.
- Pradyumna Chari (MIT). No RFCs. This is his only active draft.
All three are new to the IETF. Section 14 says Project NANDA operates NandaIndex.
Questions for the Working Group
- Does the WG want a non-DNS protocol spec now, next to the DNS candidates, or after they settle?
- Is this a separate protocol or a lookup-by-identifier profile of an ARD registry?
- Can DAWN define Authority Evidence for email, DID and workload identifiers without doing identity management, which the charter excludes?
- Should a DNS-based referral to a registry, which the charter permits, replace static configuration?