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

Review · November 16, 2026

BEST, an Agent Interaction Protocol in a Discovery Slot

draft-dinuzzo-best-protocol -03View on Datatracker →

The Problem

An AI agent that has never seen a service needs to be told, out of band, which operations it accepts, what the payloads look like, and how to get a credential.

BEST wants an agent to get from a domain name to working use by itself. The person says "use example.com". The agent fetches a manifest, reads the catalogue of operations, signs the person in, and starts sending commands (Section 1).

How It Fits In

This is not a DAWN draft by name, and the text never mentions DAWN. It has a 5 minute slot in the protocol section of the DAWN session.

So is it about agent discovery? A small part of it is. BEST is a full application protocol with a message envelope, commands, events, queries, workflows, an MCP binding, and an OAuth profile. Discovery is Section 3, about 9 of 47 pages.

The discovery part does fit the charter's definition. The DAWN charter limits discovery to a client "requesting and obtaining the minimum public discovery information for an AI resource". It excludes "open-ended search, semantic matchmaking, ranking, or selection". BEST draws the same line in Section 3.2: name resolution answers "where is example.com's service", not "which service makes videos". The manifest is public and carries reachability, transports, services and authentication requirements.

Most of the rest is out of the charter's scope:

  • "All communication beyond the initial discovery process" covers Sections 5 to 9.
  • The charter says "catalogs of capabilities are out of scope". That covers the command and query catalogues.
  • "Identity management of AI agents" plausibly covers agent registration (Section 4.3).

Which charter deliverable does it address? Only the Protocol deliverable, and only Section 3.

Section 3 still has gaps against the charter:

  • The charter names three deployment contexts. BEST serves only the public Internet one. Section 3.2 tells consumers to refuse a name that lands on a loopback, link-local or private address outside development use.
  • The charter says the working group will consider DNS and mDNS. BEST uses DNS only as a namespace. There are no new records, unlike DAN. Everything is an HTTPS fetch of /.well-known/best.
  • The charter requires integrity of all discovery information. BEST relies on TLS. Section 3.4 mentions "integrity data covering the manifest" as a possible vendor extension and defines nothing.

Summary

  • Manifest (Section 3): a JSON document at /.well-known/best, always public.
  • Name resolution (Section 3.2): fetch https://N/.well-known/best and follow redirects. The agent must only resolve names the person gave or confirmed.
  • Commands, events, queries (Sections 5 to 7): POST /commands takes a CloudEvents envelope. Results arrive later as events tied by correlationid.
  • Agent registration (Section 4.3): the agent runs the OAuth Device Authorization Grant (RFC 8628). The person opens the link in their own browser and approves.

Section 11 lists three production deployments, all by the author or the author's organisation, and one independent deployment.

Interoperability

Significant interoperability concerns

The OAuth profile is the part OAuth people should read. Section 4.3 says RFC 8628 is used "unchanged", then adds three request parameters, three token response members, and a "credential block" with seven more. None of these are registered. Section 13 says the document has no IANA actions. token_type can be apikey, which an OAuth client library will not know what to do with.

client_id means nothing. Services "pre-register no clients" and the value is "caller-declared and, like source, never identity". The approver sees agent_label, which the caller also writes.

Device code phishing is not discussed. An attacker can start the flow and send the victim the link. The victim approves on the real site and the attacker's poll gets a durable credential. Section 12 covers the agent opening the link itself, not an attacker handing the link to a person. RFC 10027 is the OAuth working group's guidance on this and is not cited.

Endpoint discovery duplicates existing metadata. tokenUrl and deviceAuthorizationUrl live in the BEST manifest. RFC 8414 and RFC 9728 already do this job. Neither is referenced.

The well-known name was refused. Section 13 says a provisional registration of best was declined by the designated expert in August 2026. The draft uses it anyway.

Mechanisms that could be removed

  • The ban on other descriptions. Section 3.1 says a deployment MUST NOT publish an llms.txt, a skill file, or an OpenAPI document of its BEST endpoints.
  • The verbatim sign-in guidance. Section 4.4 fixes three paragraphs of instructions addressed to a language model, and conformance item 11 requires them.
  • The "(Required from 0.10.0)" markers. They track the release plan of an outside "living specification" at version 0.9.20.

For DAWN, everything outside Section 3 could go.

Writing Quality

The prose is clear and confident. It does not read like an IETF document. Rules are stated as aphorisms: "The manifest is the only entry." Some sentences need two reads, such as "Endpoints MUST be served over HTTPS, but for the cleartext this paragraph allows."

At 47 pages it specifies five things at once: a discovery document, a messaging envelope, an RPC surface, an OAuth profile, and client behaviour rules for language models.

Authors

  • Riccardo Di Nuzzo. No affiliation listed. No published RFCs on Datatracker. This is his only draft.

Questions for the Working Group

  • Would the author split Section 3 into a short standalone draft, the only part DAWN could take as a candidate for the Protocol deliverable?
  • Is a manifest at a well-known HTTPS URI acceptable next to the DNS and mDNS proposals, given it has no answer for the local network context?
  • How would the manifest's integrity be protected beyond TLS?
  • Will the device grant profile go to the OAuth working group for review of its unregistered parameters, meaningless client_id, and missing phishing analysis?