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

Review · November 16, 2026

Agent Gateways as Discovery Proxies in DAWN

draft-miao-dawn-agw-intergration-consideration -00View on Datatracker →

The Problem

Agents behind a NAT or firewall cannot publish themselves in public DNS. Small agents may not be able to run wide-area discovery at all.

This draft puts an Agent Gateway (AGW) at the domain boundary to do discovery for them. Outbound, it looks up remote agents on behalf of local ones. Inbound, it answers for the agents behind it and hides their addresses.

How It Fits In

Which "DAWN Framework" does the title mean? Not draft-zhang-dawn-agent-discovery-framework. That draft is never cited. Section 1 describes the framework as a lifecycle of identity provisioning, authentication, authorization, a discovery substrate access point, invocation and monitoring. That is Figure 1 of the problem statement, the only DAWN draft in the references.

The two drafts overlap heavily and do not line up. The Zhang framework already has a Federation Gateway that applies export policy, publishes permitted information in DNS, resolves remote domains and validates the results. The AGW is the same box under a different name, with a different DNS-SD service type. This draft uses _dawn-agw._tcp. Zhang's example uses _agent._tcp. The mDNS draft uses _agents._tcp.

Which charter deliverable does it address? None cleanly. It is Informational and defines no wire format, so it is not the Protocol deliverable. It is closest to input for the Discovery Architecture document, which covers deployment models. Its three use cases (Section 8) could feed the Use Cases and Applicability document.

Section 5 argues the draft is in scope. Much of that holds. The charter says an AI resource may be "a service such as a catalog that will further the discovery process", and that discovery may use referrals and indirection. A gateway that answers for a domain fits.

Section 5 misses one exclusion. The charter says DAWN "does not include open-ended search, semantic matchmaking, ranking, or selection of an AI resource based on capabilities or user intent". The AGW's Capability Matcher does that. It matches skill IDs, version constraints, load and available capacity (Sections 3.2 and 3.3).

Mode A also has the AGW query "public DAWN indexers" (Section 3.1). The charter excludes "AI resource indexing and searching across the wider Internet".

Summary

Mode A, outbound (Section 3.1). A local agent sends POST /local-discovery to its AGW with a skill name. The AGW queries DNS-SD for _dawn-agw._tcp in the remote domain, fetches the remote gateway's /.well-known/agent.json, checks certificates and policy, and returns a filtered list.

Mode B, inbound (Section 3.2). Sub-agents register their skills with the AGW. An external agent fetches /.well-known/agent.json from the AGW. The AGW matches the request against its registry and returns an "aggregated MDI" with its own public URI as the endpoint.

DNS roles (Section 4). The AGW is the mDNS responder for its sub-agents and the DNS-SD publisher for the domain. Stable attributes go in TXT records with 120 to 300 second TTLs. Load and availability are served over HTTPS only.

Interoperability

Significant interoperability concerns

Nothing is specified to the point of implementation. POST /local-discovery has no request or response format. The TXT keys are not listed. The "DAWN MDI Schema" in Table 2 does not exist yet.

The inbound query has no query. Mode B is a plain GET /.well-known/agent.json. The AGW then matches skill=hvac_control. The draft never says how the skill gets into the request.

Aggregated MDI conflicts with the MDI model. draft-cui-dawn-mdi-model -02 Section 8.5 says an MDI instance describes exactly one entity and aggregation is a carrier matter. This draft's MDI describes a whole domain.

Mechanisms that could be removed

  • Human-in-the-loop approval (Sections 3.4.3 and 7). Section 5.1 lists it as out of scope for DAWN, then Section 7 puts a MUST on it.
  • The A2A Gateway, Network Gateway and STUN/TURN terms in Section 2. They are never used for discovery.

Writing Quality

The draft reads like a product architecture document. The AGW is "essential universal infrastructure" and discovery completes "in < 10 ms" (Section 8.2) with no basis given.

  • It is Informational but uses MUST throughout Section 7, with no requirements language section.
  • IANA Considerations is "TBC". The draft uses a new service name and a well-known URI.
  • Section 4 says sub-agents often lack mDNS stacks. Section 8.1 has a thermostat registering over mDNS.

Authors

From their Datatracker profiles:

  • Chuanyang Miao (ZTE). No RFCs. 7 active drafts, mostly in CATS and routing.
  • Ran Chen (ZTE). 5 RFCs, all routing. 37 active drafts. Secretary of the Routing Area Directorate.

Questions for the Working Group

  • How does the AGW differ from the Federation Gateway in the Zhang framework?
  • Does the Architecture document need to cover a gateway that answers for hidden agents?
  • Is the gateway's capability matching the "matchmaking" and "selection" the charter excludes?
  • Can the working group settle on one DNS-SD service type before a fourth one appears?