Review · November 16, 2026
ARD Initial Contact: Finding the Agent Manifest
The Problem
ARD puts an organization's agents, MCP servers and skills in one JSON manifest. A client needs the URL of that manifest before it can do anything else.
This draft defines how the client gets that URL from a domain name, from the local network, or from configuration. It stops there. "Initial contact ends at the manifest" (Section 3).
How It Fits In
This is one of three ARD documents sharing a 10 minute slot on Thursday. The architecture draft describes the roles. The manifest and lookup draft defines the format, the search interface, and the signature that binds an entry to its publisher. This draft is the smallest and the most finished of the three.
It maps to the charter's Protocol deliverable, and only to the first step of it. It does not define the minimum public discovery information that the deliverable also calls for.
Is finding a catalog in scope? Yes. The charter says the AI resource "may be an individual named agent or a service such as a catalog that will further the discovery process". The mechanisms cover the charter's three deployment contexts, and use DNS and mDNS as the charter suggests.
Two charter lines are harder to meet. The charter says work items "will provide integrity of all the discovery information exchanged". This draft makes DNSSEC a SHOULD and hands publisher binding to the manifest draft (Section 6.1). The charter also puts "catalogs of capabilities" out of scope. This draft only locates the manifest, so that question lands on the manifest draft. But the two are only useful together.
Summary
Every mechanism produces "entry points". An entry point is an https URI plus a record of where the client learned it (Section 3).
- Well-known URI (Section 4.3.1):
https://<domain>/.well-known/ard.json. Mandatory for publishers and clients. One redirect allowed. - Index record (Section 4.2.1): an SVCB record at
_index._agents.<domain>. The TargetName is a host that serves the manifest at the well-known path. - DNS-SD (Section 4.2.2): PTR, SRV and TXT records under
_ard._tcp.<domain>. Apathkey in the TXT record can point somewhere other than the well-known path. - Link relation (Section 4.3.2): a
Link: <...>; rel="ard"header on a HEAD request to the root. A fallback only. - Local network (Section 4.4): a domain name learned from DHCP or Router Advertisements, or an mDNS browse of
_ard._tcp.local..
Initial contact gives you a manifest from a TLS-authenticated host and nothing more. It "establishes nothing about the publisher of any entry" (Section 6).
So how does this compare to the DNS proposals? DAN and DNS-AID put per-agent data in DNS. DAN defines new record types and requires DNSSEC. This draft puts one pointer in DNS and keeps everything else in HTTPS. It needs no new record types.
Interoperability
The well-known URI is mandatory on both sides. That is the right call. Two independent implementations will find each other.
Significant interoperability concerns
Two drafts define the same DNS name. DNS-AID owns _index._agents.<domain> and says the data behind it is out of scope. Its own examples add a well-known= parameter, with ard.json in one and agents-index.json in the other. This draft hardcodes /.well-known/ard.json and ignores unknown keys. An ARD client will probe the wrong path on the second example. DNS-AID is only an informative reference.
The three ARD documents disagree with each other. The manifest draft says initial contact uses _entries._agents and _search._agents (its Section 4.1). This draft uses _index._agents and lists the difference as open issue 2. The architecture draft says initial contact locates a manifest "or an ARD Registry". This draft yields manifests only (Appendix B, item 6).
The squatting defense rests on an optional field. Section 6.4 says a tenant's entries "will not pass the check" in the manifest draft. That check needs a trustManifest, which the manifest draft marks OPTIONAL. Its own example manifest has none.
mDNS mostly does not work without public DNS. Names under .local cannot get public certificates. The publisher SHOULD use a unicast name with a real certificate, or the outcome is "not permitted" unless a key is pinned (Section 4.4.2). Zero-Configuration Agent Discovery builds https://room1.local URLs and has the same problem. The two drafts register two different service types for the same link.
Mechanisms that could be removed
- DNS-SD under
_ard._tcp.<domain>for unicast DNS. Itspathkey lets it return a different manifest than the well-known URI (Section 4.2.2), so clients that skip it see a different set of agents. - The link relation, or the well-known suffix, if open issue 1 moves ARD under RFC 9727
api-catalog.
Writing Quality
This is a well written draft. It uses standard IETF language, cites section numbers precisely, and explains what it rejected and why. Appendix B lists six honest open issues.
Two references are stale. It calls the mDNS draft's service type _agent._tcp, which revision -01 of that draft renamed to _agents._tcp. It cites DNS-AID under its old DNSOP name.
Authors
From their Datatracker profiles:
- Jaime Jimenez (Ericsson). 2 RFCs. 11 active drafts, including all three ARD documents. Chair of CoRE.
- Junjie Bu (Google). No RFCs. 3 active drafts, all ARD.
- Jari Arkko (Ericsson). 58 RFCs. 3 active drafts. Chair of PROCON.
Questions for the Working Group
So what should DAWN decide on Thursday?
- Does a pointer with optional DNSSEC and an optional publisher signature meet the charter's integrity requirement?
- Who owns
_index._agents, and what parameters does the record carry? - Does initial contact end at a manifest, or can it return a registry endpoint?
- Does DAWN want one mDNS service type or two?