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

Review · November 16, 2026

A Two-Layer Agent Discovery Framework for DAWN

draft-zhang-dawn-agent-discovery-framework -09View on Datatracker →

The Problem

DAWN has many discovery proposals and no agreed picture of how they fit together. This draft offers that picture: local discovery inside a domain, DNS between domains, and a gateway in each domain that decides what leaves.

How It Fits In

This is an individual draft aimed at the charter's Discovery Architecture deliverable: informational documents "describing deployment models, requirements (including the minimum public discovery information), and the resulting DAWN architecture". It defines no protocol, no wire format, and no IANA registrations (Sections 1.2, 4.5, 10). It is not a candidate for the Protocol deliverable.

It is one of several architecture candidates. draft-yao-dawn-agent-discovery-architect and draft-bu-dawn-ard-architecture propose different ones.

The -09 revision is close to a new document. In -08, the Federation Plane had BGP-like structured peering, gossip dissemination, and a TLV-encoded record. In -09, DNS publication and on-demand resolution are the baseline. Peering and dissemination are now "additional functions" with no protocol (Section 7). That matches the charter, which says the working group "will consider DNS and mDNS as mechanisms upon which to build a discovery protocol".

The scope tracks the charter closely. Section 1.2 repeats the charter's out-of-scope list nearly word for word. Referral to a catalog or another discovery endpoint (F-8) is in scope. The charter says discovery "may use referrals, direct transfer, delegation, or indirection".

One thing to watch. The charter says this deliverable "may be held until the Protocol work is considered stable". An architecture written before the protocol is chosen may need another rewrite.

Summary

Local Discovery Plane (Section 6.1). Agents advertise by mDNS/DNS-SD, or infrastructure such as an orchestrator supplies the records. No local protocol is mandated.

Federation Plane (Section 6.2). Each domain publishes approved records in its own authoritative DNS zone. Other domains resolve them when asked. Nothing is pushed or replicated.

The Federation Gateway (FGW) sits between the layers. Its Export Policy Engine picks which agents and fields may leave the domain. On the requesting side, the FGW checks its cache, does a DNSSEC-validated lookup on a miss, applies import policy, and caches the answer for no longer than the remaining TTL (Section 6.3.2).

Security is delegated to DNSSEC. There is no record signature, no federation membership, and no admission protocol (Sections 8.2, 8.4).

Does It Fit the Protocol Proposals?

DNS-AID: mostly. draft-mozleywilliams-dawn-dnsaid publishes SVCB records in the organization's own zone, which is the baseline. Two mismatches. DNS-AID clients query DNS themselves, while this framework routes cross-domain requests through an FGW. DNS-AID says records SHOULD be DNSSEC-signed and is "deployable without DNSSEC". The framework's integrity requirement (F-9, NF-1) rests entirely on DNSSEC. The charter says work items "will provide integrity of all the discovery information exchanged", so this gap matters.

DAN: yes, but uncited. draft-seethiraju-dawn-dan requires DNSSEC validation and defines its own record types. It fits the baseline cleanly. The framework does not mention it.

ARD: only in the undefined part. draft-bu-dawn-ard-architecture is built on registries, manifests, an HTTP query interface, and semantic search. The framework puts capability catalogs and semantic matchmaking out of scope (Section 1.2), and so does the charter. The charter does allow "a service such as a catalog that will further the discovery process" as the thing discovered. So ARD lands in the "trusted directory" and "referral" boxes of Section 7, where every protocol detail is out of scope. The framework does not cite ARD either.

MDI: out of date. Section 1 says MDI has three mandatory elements: Entity Identifier, Entity Type, Endpoint. That is MDI -00. In -02, the mandatory elements are Name and Endpoint, and Entity Type is optional. MDI -02 in turn describes this framework as defining a TLV binding, which -09 removed.

Interoperability

Significant interoperability concerns

The FGW has no interface. Section 6.2 says a client asks the local FGW for cross-domain information. No protocol for that request is named. If the client can query DNS directly, as in DNS-AID and DAN, the requesting-side FGW is a caching resolver with an import filter.

The baseline does not meet the draft's own requirement. F-3 says each site MUST control sharing with "per-peer access policy". Section 8.3 admits public DNS "does not provide arbitrary per-user disclosure control". Per-recipient sharing moves to Section 7.1, with no protocol.

Requirements use MUST without force. Section 3 calls them "non-normative". F-2 says agents MUST be able to self-advertise, and Section 6.1.1 makes it a MAY.

Mechanisms that could be removed

  • Section 7. It names three functions and scopes out everything about them.
  • Sections 9.1 to 9.4. They restate DNS TTL and caching behavior.

Writing Quality

The prose is readable. At 37 pages it is long for what it specifies. The rewrite left pieces of -08 behind:

  • Sections 1.2 and 4.3 cite "Section 7.2.3". It does not exist.
  • Section 9.6 says to monitor "peer session uptime" and "signature verification failure rate". The baseline has no sessions and no signatures.
  • Section 4.4 puts DNS in the Local Discovery Plane, and Section 6.2 makes it the Federation Plane.

Authors

From their Datatracker profiles:

  • Bin Zhang (Pengcheng Laboratory). No RFCs. 11 active drafts, mostly in CATS and DAWN.
  • Linda Dunbar (Futurewei). 10 RFCs. 11 active drafts, mostly routing and SD-WAN.
  • Yixuan Yang (Sun Yat-sen University). No RFCs. 3 active drafts.

Questions for the Working Group

  • Is a Federation Gateway part of the DAWN architecture at all, and if so, what protocol does a client use to ask it?
  • Is DNSSEC required, given that the charter requires integrity and DNS-AID says SHOULD?
  • Is a referral to an ARD registry all the architecture needs to say about ARD?
  • Should architecture adoption wait until the protocol choice is made, as the charter permits?