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

Review · November 21, 2026

Agent Identity and Discovery (AID) with a DNS TXT Record

draft-nemethi-dawn-aid -00View on Datatracker →

The Problem

A user types a domain into an agent client. The client needs the agent's endpoint and the protocol to speak there. Today that comes from configuration or a directory.

AID answers with one DNS TXT record at _agent.<domain>. The record carries a URI, a protocol token such as mcp or a2a, an authentication hint, and an optional public key the endpoint can prove it holds.

How It Fits In

This is a candidate for the DAWN charter's Protocol deliverable. The charter names DNS as a mechanism to build on, and AID's "where is it and what does it speak" is close to the charter's "minimum public discovery information".

It is a -00 under this name, but not new work. It replaces the expired draft-nemethi-aid-agent-identity-discovery and describes a second wire format, aid2. The draft says AID has been deployed since mid-2025 with SDKs in six languages.

Three charter lines are a harder fit:

  • One agent per domain. Two valid aid2 records at the same name is an error (Section 4.1). The charter's public Internet case is "discover all agents of example.org". AID has no index and no way to list.
  • Integrity. Work items "will provide integrity of all the discovery information exchanged". DNSSEC is a SHOULD on both sides (Section 5).
  • Identity. The charter excludes identity management. Despite the name, AID only proves an endpoint holds a key. The draft says so itself in Section 5.1.

So how does it differ from the other DNS candidates?

  • DNS-AID uses SVCB records, one per agent name, plus an _index._agents pointer to a registry. AID uses a single TXT record and has no index.
  • DAN defines new record types and makes DNSSEC a MUST.
  • ARD initial contact considered AID's approach and cites it as prior art.

AID depends on no OAuth RFCs. It points to them: after discovery, the client follows Protected Resource Metadata at the endpoint (Composition Notes).

Summary

The domain owner publishes one record:

v=aid2;u=https://api.example.com/mcp;p=mcp;a=oauth2_code;k=<Ed25519 key>

The client queries the exact host the user entered. It never walks up to a parent domain. An operator who wants sharing uses a CNAME (Section 4.3).

If k is present, the client must run an endpoint proof called PKA. It sends a nonce to the endpoint. The endpoint returns an HTTP Message Signature (RFC 9421) over the method, target URI, authority, status, and nonce (Appendix A).

Domain binding is an optional add-on. The client sends an AID-Domain header with the domain it queried, and the endpoint signs that too. This shows the endpoint agrees to be that domain's agent, since anyone can publish a record pointing at someone else's endpoint and key (Section 5.1).

If DNS returns nothing, the client may fetch https://<domain>/.well-known/agent instead (Appendix B).

Interoperability

Significant interoperability concerns

The PKA request is not defined. Appendix A.3 lists headers. It never says which HTTP method the client sends, or whether there is a body. The signature covers @method, so both sides must agree. A GET to an MCP endpoint and a POST get different answers.

The proof key comes from the same record it protects. Without DNSSEC, an attacker who spoofs the TXT record replaces u and k together, and PKA verifies. With honest DNS, TLS already authenticates the host in u. The draft needs to say what PKA adds over TLS.

Domain binding does nothing by default. In the default prefer mode an unbound proof is accepted. Section A.7 admits an attacker's endpoint can ignore the header. Only require mode closes the gap, and that is a local policy choice.

The .well-known fallback is six bullet points. The format is "JSON mirroring v2 record keys". There is no schema, no rule for long keys versus aliases, and no IANA request for the agent well-known URI.

aid1 is required but not defined. Clients must parse, partition, and downgrade-check aid1 records. The format lives only in an expired draft cited as an informative reference.

The token registries are a GitHub repository. The auth and proto lists are maintained outside IANA (Section 9). Tokens such as ucp, zeroconf, and custom have no definition or reference.

The auth hint duplicates OAuth discovery. a=oauth2_code says less than Protected Resource Metadata, and the draft tells clients to follow that anyway. If the two disagree, the draft gives no rule.

Mechanisms that could be removed

  • proto=local. A TXT record that tells the client to run docker:, npx:, or pip: packages. This is software installation, not discovery of a reachable resource. It also brings five client safeguards with it (Section 5).
  • Enterprise policy presets (Section 5.3) and numeric error codes (Section 4.2). These are SDK design. Neither affects what goes over the wire.
  • The agent service name request (Section 10.2). It reserves a name for "possible future SRV-based discovery". Nothing in the draft uses it.
  • Protocol-specific names like _agent._mcp.<domain>. They are described as legacy and diagnostic only.
  • Sections 7.1 and 7.2, Composition Notes, Migration Notes. These explain what AID does not do. One paragraph would cover it.

Writing Quality

The core is clear. A reader can understand the record format and lookup in two pages. Section 5.1, "What PKA Proves", is unusually honest about limits.

The rest reads like project documentation converted to a draft:

  • A third of the text explains non-goals and design rationale.
  • The cross-origin redirect rule is stated twice in Section 5.
  • The prior draft is labeled [RFC-NEMETHI-AID-00]. It is not an RFC.
  • A note to editors is left in: "Specification and reviewer-facing language should say...".
  • Normative client behavior sits in appendices, and the last two appendices are unnumbered.

Authors

From the Datatracker profile:

  • Balazs Nemethi (Open Agent Registry, Inc.). No RFCs. 1 active draft, this one, and 1 expired draft, its predecessor. A new IETF author.

Questions for the Working Group

So what should DAWN ask of this draft?

  • Is one agent per domain enough, when the charter asks for "all agents of example.org"?
  • TXT or SVCB? DNS-AID carries the same endpoint and protocol data in SVCB.
  • Who gets the label? AID requests _agent, DNS-AID uses _agents, and the mDNS draft just moved from _agent._tcp to _agents._tcp.
  • Does PKA belong in a discovery protocol, or is DNSSEC plus TLS the integrity story?
  • Should the auth and proto tokens move to IANA and align with the common parameters draft?