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

Review · November 22, 2026

Generalizing the DAWN MDI for All Entity Types

draft-zhangb-dawn-entity-mdi -00View on Datatracker →

The Problem

DAWN drafts are starting to define a separate discovery mechanism for each kind of thing. Agents get _agents._tcp. Local LLM APIs get _llm-api._tcp with their own TXT keys. Tools, models, and data sources would each need another.

This draft says to stop. It proposes one DNS-SD service type for everything, with a new rtype key that says what kind of entity each instance is.

How It Fits In

This is a layer on top of three other individual drafts. None of them are adopted.

It adds one TXT key, two IANA registries, and ten entity type values. It has no OAuth dependency.

Does it fit the charter? Partly. The charter covers "AI resources" and names catalogs as one kind. Agents, catalogs, and services are easy to justify. Tasks, workloads, brokers, and application endpoints are a stretch for a group chartered to do initial discovery of public information.

The bigger problem is that the draft is built on old versions of its dependencies.

The MDI model it inherits no longer exists. The draft copies the element set from -00 of the MDI model: three mandatory elements, plus Capability Summary, Capability Reference, Trust Reference, Provenance, Scope Hint, and Operational Hint. The MDI model dropped most of those by -02, published before this draft. The current model has Name and Endpoint as mandatory, and Entity Type is a MAY. This draft says the element set is "exactly that of" the MDI model. It isn't.

The mDNS keys it reuses were removed. The draft lists agentId, cardUrl, caps, and ver as keys from the mDNS draft. Revision -01 dropped all four in October. It also renamed the service type to _agents._tcp, so the "two spellings to be reconciled" issue this draft raises twice is already settled.

The auth key is not a common key. The draft says auth is already in the common parameters draft. That draft defines nine keys, and auth is not one of them.

How It Works

Every entity is advertised under _agents._tcp. Two keys describe it.

  • rtype says what it is: agent, catalog, service, tool, task, data-source, model, workload, broker, or application-endpoint. Absent means agent.
  • proto says how to talk to it: a2a, mcp, openresponses, or https.

The two vary independently. An MCP server is proto=mcp with rtype=tool or rtype=agent. A local LLM API is proto=openresponses with rtype=service.

A client browses once and gets everything on the link. It then fetches the native description, such as an AgentCard or model card, and continues in the entity's own protocol.

New entity types are added with a registry entry and a profile document. Only two profiles exist: agents, which is the existing mDNS draft unchanged, and LLM API services, which is a key-by-key remapping of the _llm-api._tcp draft.

Catalogs get special handling. Fetching a catalog's description returns a list of more entities, and the client repeats the process on each one, up to a suggested depth of 3.

Interoperability

The core argument is right. One service type with a proto key is where the DNS drafts already landed, and a second service type for LLM APIs repeats work. That part needs no new draft though. It needs openresponses as a proto value and one paragraph in the common parameters draft.

Significant interoperability concerns

rtype changes nothing a client does. The draft says Entity Type "does not alter the semantics of any other element" and clients MUST ignore values they don't recognize. A client that sees proto=mcp connects with MCP whether the instance says tool or agent. A key with no effect on behavior is not needed for interoperability.

The one type that does change behavior is unspecified. A catalog triggers recursive fetching. The catalog document is "a list whose entries are MDI instances or references to them." There is no format, no media type, and no retrieval rule. Appendix A says no encoding is normative. Two implementations cannot exchange a catalog.

The same thing can be advertised under different types. An MCP server is a tool or an agent "according to its role." An LLM endpoint fits both service and model. Publishers will choose differently. A client browsing _tool._sub._agents._tcp misses the MCP servers that picked agent.

Type and protocol subtypes share one namespace. The draft uses _openresponses._sub for a protocol and _tool._sub for a type. Both live under _agents._tcp. Nothing stops a protocol from registering a name that matches a type.

The draft contradicts itself on the URL scheme. Section 6.2 says https unless a binding says otherwise. Appendix C says the scheme is derived from alpn. Plain HTTP local servers, which is most of Ollama and LM Studio today, have no way to say so.

https is registered as a proto value. The common parameters draft keeps HTTP versions in alpn and the application protocol in proto. A proto=https instance tells the client nothing about what to send.

Recursion is SHOULD in one section and MUST in another. Section 7.3 says catalog recursion SHOULD be bounded. Section 10 says it MUST be.

Scope says no DNS binding, then defines one. Section 3 says the draft "does not bind MDI to DNS or to any other carrier." Section 7 is titled "Discovery Procedure and DNS-SD Binding" and defines a TXT key.

Mechanisms that could be removed

  • Seven of the ten entity types. task, workload, broker, data-source, model, tool, and application-endpoint have a one-line definition and no profile. Their protocols are listed as "the protocol registered for the deployment." Register them when someone has a use for them.
  • The restated information model. Sections 4.1 and 4.2 and Table 1 copy another draft and say "unchanged" thirteen times. The copy is already out of date.
  • The DAWN Entity Protocols registry. A proto registry belongs in the common parameters draft, which both DNS drafts already point to for it.
  • rtype as a TXT key. The common parameters draft already explains why subtypes beat TXT keys for coarse categories: a client has to receive every TXT record before it can filter on a key.
  • Appendices B and C. Both map against key names that no longer exist.

Writing Quality

The prose is fluent and well organized. RFC 2119 language is mostly used correctly.

It is long for what it adds. Twenty-five pages deliver one TXT key and a list of ten words. The motivation section makes its point five times in a bulleted list.

There are signs of a fast draft. The new constraint is numbered "1. 6." The phrase "no new service type is needed for a new entity type, ever" is not typical spec language. Claims like "No information is lost" and "adds no new attack surface" are stated without support.

The security section is thin for what the draft does. One browse now reveals every model, data source, and workload on the link. The mitigation is to advertise less.

Background of the Author

Bin Zhang (Pengcheng Laboratory) has no RFCs. He has 11 active drafts and 2 expired ones, so he is writing a lot right now.

Seven of the active drafts are in CATS, on computing-aware traffic steering. Four are in DAWN: this one, the agent discovery framework, BGP-style federated discovery, and infrastructure-assisted discovery as a co-author.

Questions Worth Raising

  1. What does a client do differently based on rtype? If the answer is nothing except for catalog, is a ten-value registry needed?
  2. Will this be rebased on the current MDI model and mDNS draft? Most of the inherited elements and keys are gone.
  3. Why a separate draft? Could "LLM APIs are one more proto value" be a change to the common parameters draft?
  4. What is the catalog document format? Recursive discovery cannot work without one.
  5. Who decides whether an MCP server is a tool or an agent?