Review · November 22, 2026
The Intent-to-Capability Knowledge Gap in DAWN Summary
The Problem
DAWN discovery assumes the client already knows what to ask for. This draft says that assumption is wrong for agents.
An agent starts with an intent, such as "get to place X as fast as possible". It doesn't start with a capability name. The right capabilities depend on what is available nearby: a bike share, a taxi, or a train plus a flight. The client only learns what is available by running discovery, and it can only run discovery once it has picked capabilities to ask about.
That is the circular dependency. The draft names it, shows where the current DAWN documents skip over it, and compares five ways to handle it.
The author separates this from two nearby problems. It isn't bootstrapping, because the client is assumed to have an entry point. It isn't vocabulary definition either. A perfect dictionary tells the client which names might exist, and says nothing about which are offered in this scope.
How It Fits In
This is a discussion contribution. It defines no protocol, wire format, or registry, and has no IANA actions.
Does it fit the charter? The document does, as input to the architecture deliverable. The topic sits right on the charter's boundary. The charter excludes "semantic matchmaking, ranking, or selection of an AI resource based on capabilities or user intent." The draft stays on the right side of that line by asking a narrower question: what does the minimum discoverable information (MDI) have to contain so the client can do the matching itself?
It adds four terms on top of the terminology draft: Intent, Decomposition, Capability Identifier, and Scope Enumeration. It creates no new roles.
The dependencies are six individual drafts and RFC 6763 (DNS-SD). Nothing it cites is an adopted DAWN document, because DAWN has none yet. There is no OAuth or WIMSE dependency.
High-Level Summary
The draft compares five strategies. They differ in who does the matching and who sees the intent.
- A. Identifier probing. The client holds a vocabulary and queries for each candidate capability. This is the DNS-SD model. It needs the full vocabulary in advance, and the pattern of probes gives the intent away.
- B. Intent forwarding. The client sends the intent to a catalog that does the matching. No vocabulary is needed. The intent is disclosed, and this is the semantic matchmaking the charter excludes.
- C. Scope enumeration. The client asks for everything in a scope and matches locally. The intent never leaves the client. It only works for small scopes, and only if the MDI is descriptive enough to interpret.
- D. Brokered discovery. The client discovers a broker, such as a journey planner, by a coarse identifier. It hands over the intent later, in the application exchange. The broker learns everything.
- E. Reversed discovery. The client publishes its need and providers find it. Everyone who can see the scope learns the intent.
The draft concludes that A, C, and D fit the charter as written. It suggests a layered combination: a coarse probe finds the scope or broker, enumeration runs inside a small scope, and a broker handles intents that cross scopes.
The scalability section makes a good point about option C. For an LLM-based client, the limit on enumeration isn't bandwidth. It is context budget and the cost of interpreting every entry.
The security section follows from that. Descriptive MDI read by a language model is untrusted input. A publisher can write a description to match as many intents as possible, or to steer the client's behavior. Signing the record proves it wasn't altered. It doesn't prove the description is honest or safe to read.
The draft closes with five questions for the architecture work: what the MDI contains, which scopes allow enumeration, what a client does with a truncated result, how intents that span several scopes get resolved, and who owns the capability vocabulary.
Is This Document Necessary?
There is nothing to implement, so interoperability isn't the right test. The question is whether the working group needs this pointed out.
It does. The gap is real, and the draft finds two tensions in the requirements that should be settled before the MDI is fixed.
Enumeration versus scraping. The charter's own examples are all enumerations: "discover all agents in this shop." The requirements draft also requires limiting scraping (REQ-SEC-3), and says a mechanism that doesn't resist enumeration should treat its entity list as public. The architecture has to say which scopes allow enumeration, and to whom.
What "services it offers" means. REQ-PROP-1 requires the MDI to include services, and doesn't say whether that is an identifier or a description. Option A needs an identifier and a registry behind it. Option C needs a description. The MDI draft already made a choice here. It has no separate Services element, and it puts coarse capability labels beyond the minimum. Under that model, option C mostly doesn't work.
That is the main thing missing from this draft. It reviewed the charter, requirements, terminology, use cases, problem statement, and privacy drafts. It did not review the MDI draft, which is where its first question is being answered. The next revision should engage with it directly.
The draft also stops short of a recommendation. It says which combination "scales best on paper" and then lists open questions. Its privacy section does take one position: keep the intent out of the discovery exchange, and carry it only in an authenticated application exchange with a chosen party. That points at option D, and the draft would be more useful if it said so.
Option D is also how this works today. A user asks a travel site or an assistant with a tool catalog, and that party does the planning. If brokers are the practical answer, the gap is handled above DAWN, and DAWN's job reduces to discovering a broker by a coarse type. The working group should decide whether it agrees with that before spending time on descriptive MDI.
Does this need to be an RFC? No. It is a good mailing list contribution and a good set of issues against the requirements, use cases, and MDI drafts. The request to add chained and interactive discovery to the use cases draft is the most concrete of these.
One thing to watch: the citations are already stale. The draft reviews terminology -05 and use cases -01. Those are now at -07 and -02.
Writing Quality
The writing is very good. It is about eight pages, with short sentences and plain words. It has no BCP 14 keywords, which is correct for a document with no normative content.
The two comparison tables carry most of the argument and are easy to scan. The single travel example runs through the whole document and is worked through all five strategies in the appendix. The "What This Is Not" section heads off the two most likely misreadings early.
The length is right. The editorial note listing exactly which draft versions were reviewed, and which were only seen in excerpts, is a practice more authors should copy.
Two small problems. The author's email is a literal "[EMAIL TBD]" placeholder. The title uses "Discovery of Agents With Names" while the cited problem statement still expands DAWN the older way, which is the working group's inconsistency more than the author's.
Author Background
Artur Hecker (Huawei Technologies Duesseldorf) has no published RFCs. This is his only active Internet-Draft. He has five expired drafts across ANIMA, SFC, and energy management, including draft-ietf-anima-grasp-distribution, which was adopted by the ANIMA working group and later expired. He has some IETF history in network management, and this is his first draft in DAWN.