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

Review · November 21, 2026

Federated Agent Metadata Synchronization Protocol (FAMS-P)

draft-zhang-dawn-bgp-style-federated-discovery -00View on Datatracker →

The Problem

Some federations want to share different agent records with different partners, and public DNS cannot do that. This draft defines a BGP-style peering protocol where each domain's gateway pushes signed agent records to each configured peer over mutual TLS.

How It Fits In

This is an individual draft aimed at the DAWN charter's Protocol deliverable. It is a complete new binary protocol: five message types, a session state machine, a TLV record format, three new IANA registries, and an ALPN ID.

That is a hard fit for the charter. The charter asks for a mechanism that "builds on existing protocols and tools" and says the working group "will consider DNS and mDNS as mechanisms upon which to build a discovery protocol". FAMS-P builds on TCP and TLS only. The deployment context does match: "between cooperating parties with trust relationships over the public Internet".

The draft depends on an architecture that has since moved. It calls itself the realization of "Mode B" in draft-zhang-dawn-agent-discovery-framework, citing -05. The current -09 has no Mode A or Mode B. It mentions structured peering once (Section 7.1) as "one possible realization" and puts the protocol out of scope. The framework still defines the Federation Gateway (FGW), Federation Metadata Record (FMR), and Federated Agent Directory (FAD) that this draft uses.

The MDI mapping is also stale. Section 3.6 describes MDI -00, with three mandatory elements. MDI -03 has two, Name and Endpoint. MDI -03 calls out this draft directly: it "makes the Capability-Summary mandatory", while MDI places capability labels beyond the minimum. The charter puts capability information beyond the minimum out of scope.

There is a direct competitor. Section 3.1 compares FAMS-P to AMSP (draft-liu-agent-metadata-sync-protocol), says operators SHOULD NOT run both between the same gateways, and encourages consolidation.

It depends on no OAuth or WIMSE work. It uses JWS (RFC 7515) for record signatures and X.509 for gateway identity.

Summary

Each domain runs a Federation Gateway. Administrators configure peerings by hand. There is no peer discovery and no admission protocol.

Two gateways open a TCP connection with mutual TLS 1.3 and exchange OPEN messages. Each then sends its full set of records for that peer (FULL_SYNC). After that, each sends one UPDATE per agent change, either ANNOUNCE or WITHDRAW. KEEPALIVE and a hold timer detect dead peers.

Each record (FMR) describes one agent: identifier, type, endpoint URIs, capability labels, timestamp, and TTL. The originating gateway signs every record with a JWS. Receivers verify the signature before storing the record in their local directory (FAD).

Records expire unless the origin refreshes them. The default TTL is 300 seconds, with a refresh at least every half TTL.

Export policy is local and per peer. Each gateway picks which agents and fields each peer sees. Relaying a peer's records to a third gateway is off by default. The origin can permit it with a Redistribution-Scope flag.

So what does a client do? The draft does not say. It defines how gateways fill their directories and nothing about how a client queries one.

Interoperability

The session layer is specified closely enough to build. The record layer is not.

Significant interoperability concerns

Records have no boundaries. A FULL_SYNC payload is "a sequence of one or more complete FMRs". An FMR is a sequence of TLVs in any order, with no record length and no count. A receiver cannot tell where one record ends and the next begins.

The signature is underspecified. The JWS payload is the raw TLV octets. No algorithm is mandated. The key is identified "e.g., via a key identifier or the gateway's certificate reference". JWS Compact Serialization carries its payload, so every record is sent twice unless a detached form is intended. The draft does not say.

Relayed records cannot be verified without out-of-band keys. Section 9.2 says keys come from the peer certificate for first-hand records, and that key distribution is otherwise out of scope. Two implementations that both support re-distribution have no common way to verify each other's relayed records.

Per-peer redaction conflicts with the conflict rules. Export policy lets an origin send different field subsets of the same agent to different peers. Those are different signed contents under the same record key. If both copies reach one gateway through re-distribution with equal timestamps, rule 4 in Section 7.3 says to reject the record as a protocol error.

Full sync and re-distribution interact badly. On end-of-sync, the receiver replaces "all existing records previously learned from this peer". With re-distribution, the same record arrives from several peers. The draft does not say whether a record dropped from one peer's sync survives because another peer still offers it.

Passive opens are not described. The state machine covers only the side that initiates the connection. Collision resolution compares Gateway-IDs "lexicographically", and a Gateway-ID may be a DNS name or a URI.

The certificate check is loose. The certificate must carry an identifier "consistent with" the claimed Gateway-ID. No matching rule is given.

Scale. The draft targets 2 to 200 sites in a full mesh. At 200 sites that is 19,900 hand-configured sessions. Every agent is re-signed and re-sent to every peer at least every 150 seconds by default.

Mechanisms that could be removed

  • Transitive re-distribution and the Redistribution-Scope TLV. They are off by default, and most of the open problems above come from them.
  • Mandatory Capability-Summary. MDI treats it as beyond the minimum.
  • The ALPN fallback in Section 5.1, which lets a connection proceed without ALPN under local policy.
  • Section 3. Six subsections position the draft against other individual drafts and define no behavior.
  • A WITHDRAW must carry every mandatory TLV, including Endpoint and Capability-Summary. A key, timestamp, and signature would do.

Writing Quality

The draft reads like a routing protocol spec, with packet diagrams, timers, and error subcodes. The BCP 14 language is used correctly in most places. At 26 pages it is reasonably sized for what it defines.

It has not been proofread against itself:

  • The header says it expires 17 April 2027, and the boilerplate says 4 April 2027.
  • Section 5.3 points to Section 5.5 for the hold timer, which is in 5.4.
  • Section 2 points to Section 5.4 for the end-of-sync flag, which is in 7.1.
  • Sections 4, 7.4, and 8 point to the wrong sections for re-distribution, which is in 7.6.
  • Section 6.1 says TLVs "appear in any order except that, within the signed content, they are processed as an unordered set". Section 6.3 signs them in transmitted order.

The reference classification is wrong. The architecture, MDI, and terminology drafts are informative references. The draft takes its core terms from them and has a MUST "as specified in the architecture document" (Section 7.4).

The main design choice lacks justification. The draft says DNS "does not meet operational requirements" for some deployments and does not say why a new TCP protocol is needed in place of an existing replication or pub/sub mechanism over HTTPS.

Authors

From the Datatracker profile:

Questions for the Working Group

  • Does the charter leave room for a new non-DNS synchronization protocol, or is per-peer disclosure a problem to solve on top of DNS?
  • The framework no longer defines Mode B. Does this draft follow the framework, or does it now stand alone?
  • How does a client query a gateway's directory? Without that, FAMS-P replicates data and discovers nothing.
  • Should FAMS-P and AMSP merge before either asks for adoption?