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

Review · November 21, 2026

Infrastructure-Assisted Discovery for AI Agents

draft-dunbar-dawn-infra-assisted-discovery -03View on Datatracker →

The Problem

In a managed network, the infrastructure already knows two things a discovery client does not. The orchestrator knows where an agent is reachable right now. The network knows which site, tenant, VPN, slice or edge region the requester is attached to.

This draft says both should feed discovery, so the requester never has to know about UPFs, DNNs or VRFs. It names two "logical integration points" and defines neither.

How It Fits In

The draft is Informational and defines no wire protocol, no DNS-SD service type, no TXT keys and no API. The closest charter deliverable is the Discovery Architecture document, which covers deployment models.

It leans on the DAWN vocabulary. The Minimum Viable Dataset (MVD) comes from draft-farrel-dawn-terminology, where it is the data a discovery mechanism needs in order to return results. This draft's claim is that part of the MVD can come from infrastructure instead of the requester. draft-cui-dawn-mdi-model -03 already says the same thing in one sentence: other mechanisms can supply "context derived from the client's network attachment".

Other DAWN drafts already depend on it. The Zhang framework cites both hooks, and the use cases draft has a section built on them. Both authors of this draft are also authors of the framework.

Is it in charter scope? The draft works hard to say yes. It disclaims search, matchmaking, ranking and selection in almost every section, and points to CATS for network-aware selection.

One charter question is not addressed. The charter is about "public" discovery information. This draft gives different requesters different DNS views based on tenant, VPN or slice. The MDI model draft lists exactly this as an open question: is requester context part of the MVD, or is it access control outside discovery?

Summary

Resource State Hook (Section 6). The orchestrator, gateway or hosting platform tells the discovery system the agent's current external endpoint and protocol. If an agent moves from one cluster or edge site to another, the DNS records follow. Backend scaling behind a stable endpoint changes nothing.

Attachment Context Hook (Section 7). Trusted network infrastructure reduces the requester's network state to an opaque label called an Attachment Scope, such as AScope-Edge-A. Deployment policy maps that label to one or more discovery domains, such as edge-a.agent.example and shared.agent.example. The requester never sees the label.

Two-stage discovery. DAWN returns only an endpoint, a protocol and maybe a metadata link. The requester then fetches the A2A Agent Card or equivalent from the agent.

Mappings (Sections 8 and 9). For DNS-SD, resource state becomes PTR, SRV and TXT records, and the Attachment Scope picks the browsing domain or resolver view. For DNS-AID, resource state becomes SVCB records, and the Attachment Scope picks the namespace or _index._agents entry point.

Interoperability

Significant interoperability concerns

There is nothing to implement. Both hooks are "logical inputs". How the orchestrator publishes state, how the network derives a scope, how a scope maps to a domain, and how the result reaches the resolver are all "deployment-specific and outside the scope of this document". Two vendors cannot build to this and interoperate, because there is no interface between them.

The "discovery function" is never located. Section 8.3.1 says it "combines the requester-provided bootstrap information with the infrastructure-provided Attachment Scope". Is that the stub resolver, a recursive resolver, a proxy, or a client library? The requester asked for a name in example.com. Something has to turn that into a query in edge-a.agents.example.com. The draft does not say what.

The gap analysis overlooks existing DNS-SD machinery. Section 4.3 says DNS-SD defines no mapping from the requester's network state to a discovery domain. RFC 6763 Section 11 defines domain enumeration, including queries derived from the client's own IP address and from DHCP-supplied domains. Split-horizon DNS covers the "resolver view" case. The draft needs to say what these do not solve.

Naming has drifted from the documents that cite it. Revision -02 called the first hook the "Resource Publication Hook". Revision -03 renames it to "Resource State Hook". The Zhang framework -09 still uses the old name.

Mechanisms that could be removed

  • Section 10 and most of Sections 4.1, 4.5, 5.5 and 5.6. They restate the charter's bootstrap-versus-selection split and the two-stage model, which the mDNS draft and the MDI model already cover.
  • Section 6.2. A Kubernetes Service with a stable ingress is the case where nothing changes.
  • The protocol= and metadata= TXT examples. The draft says they are illustrative and defines no keys, and other DAWN drafts are defining the real ones.

Writing Quality

The sentences are clear and the terminology is consistent within the document. The problem is length. This is 48 pages for two ideas, and it grew by about a quarter since -02.

The same content appears many times. The list of UPF, DNN, slice, VPN and edge domain appears in Sections 3, 4.3, 4.4, 5.1, 5.3, 7, 7.2, 7.4 and 9.2. The discovery-versus-selection box in Section 4.5 is repeated almost word for word in Section 5.6.

  • The BCP 14 boilerplate is included, but no requirement keywords appear in the body. Section 12 uses lowercase "must" and "should".
  • IANA Considerations is "TBD". Nothing in the draft needs registration.
  • Six references are never cited in the text, including an "Agent Attachment Protocol" with no draft name. RFC 8174 is cited but not listed.
  • The DNS-AID reference points to the old dnsop draft name, and the framework reference is two revisions behind.
  • The acknowledgments thank "xxx", and Appendix A has no content apart from the author addresses.

Authors

From their Datatracker profiles:

  • Linda Dunbar (Futurewei). 10 RFCs, mostly data center, TRILL and overlay networking. 12 active drafts across SD-WAN, IDR, CATS and DAWN, and 73 expired. Reviewer for the Gen-ART, Ops, Routing and Security directorates.
  • Bin Zhang (Pengcheng Laboratory). No RFCs. 11 active drafts, nearly all in CATS and DAWN, and 2 expired. A new author writing a lot at the moment.

Questions for the Working Group

  • Should this be a few paragraphs in the Architecture document instead of a standalone draft?
  • Is information served only to one tenant or slice "public" in the charter's sense?
  • Does anything here need a protocol, or do split-horizon DNS and RFC 6763 domain enumeration already cover it?
  • If the hooks stay abstract, should the framework and use cases drafts keep citing them as if they were mechanisms?