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

Review · August 21, 2026

DNS-like Agent Discovery Architecture

draft-yao-dawn-agent-discovery-architect -00View on Datatracker →

Problem Statement

This draft proposes a three-tier hierarchical discovery architecture for the "Internet of Agents" (IoA) — an anticipated world of millions to billions of software agents that need to find each other by capability. The core problem is: given massive agent deployment scale, how do agents efficiently discover other agents that have the capabilities they need, without a single centralized bottleneck? The proposed answer is to model agent discovery after DNS, reusing DNS paradigms (hierarchical delegation, zone transfer, DNS-SD) for agent registration and capability lookup.

How the Spec Fits In

This is a DAWN (Discovering and Accessing with Networks) working group individual draft, not an OAuth or token-based security spec. It sits in the discovery layer of the agent ecosystem — beneath any identity or authorization considerations. The spec does not reference or depend on OAuth, OpenID Connect, or any authentication protocol; security is explicitly deferred ("to be added").

Within the DAWN context, this draft occupies the architectural/solution space that other DAWN drafts (requirements, use cases, gap analysis) have been framing. It is the first concrete architectural proposal to appear as an individual draft in this emerging area, and it builds on core DNS standards (RFC 1034, RFC 1035, RFC 9665) rather than OAuth-family RFCs.

High-Level Summary

The architecture defines three layers:

  • Agent Root — A globally trusted top-tier node that maintains a whitelist of valid Agent Registry suffix domains. It pushes whitelist updates to resolvers but explicitly must not store any agent business data. Think of it as the analogue to DNS root servers.

  • Agent Registry — A domain-level service where agents register. Upon registration, an agent's native FQDN (its "prefix") is combined with the registry's suffix domain to form a globally unique Composite Agent FQDN (e.g., agent-hotel.example.example-registry1.com). The registry publishes DNS records (A, SVCB, SRV) and hosts a Capability Card — a JSON document per agent describing its capabilities, network endpoint, TLS CA info, and availability.

  • Agent Resolver — A distributed discovery node that syncs the whitelist from the Root and pulls agent metadata from Registries via zone transfer. Clients query resolvers over HTTPS to search for agents by capability. Clients can find a local resolver via DNS-SD (_agent-index._tcp) or static preconfiguration.

The flow: an agent registers at a Registry → Registry publishes DNS records and a Capability Card → Resolvers sync data from Registries → client queries a Resolver by capability → Resolver returns matching agent endpoints.

Interoperability Evaluation

Significant interoperability concerns:

The biggest gap is the Agent Root governance model. The spec says there is a globally trusted root that maintains the registry whitelist, but provides no mechanism for how the root is established, who controls it, how many roots can exist, or how registries enroll with the root. Without this, implementations will make incompatible choices about trust anchors — the most critical part of the architecture.

The Capability Card format defines structure but not semantics. The capability_list field holds free-text strings capped at 200 UTF-8 characters with no defined vocabulary, ontology, or matching algorithm. Clients querying resolvers by capability will have no interoperable basis for matching unless this is specified. This is the primary value the spec offers and it is left essentially undefined.

The security section is completely absent ("to be added"). For a system built on a globally trusted root with TLS CA fingerprints in Capability Cards, security properties are fundamental to interoperability. Implementations cannot be interoperable on trust model, TLS validation, or replay/spoofing resistance without this.

Mechanisms that could be removed:

The distinction between the Agent Root "pushing" whitelist updates versus Resolvers "pulling" from registries is introduced without justification. Simplifying to a single data flow model would reduce complexity without harming the core architecture.

The agent_online_status field in the Capability Card presupposes a real-time or near-real-time update mechanism that is not described anywhere — it is either always stale or requires a synchronization protocol the spec doesn't define, and could be removed.

Writing Quality

The writing quality is below typical IETF standards for a draft intended to propose a workable architecture. Sentences are grammatically correct but frequently vague ("globally unique," "trusted whitelist," "asynchronous data replication") without the normative precision needed to evaluate or implement the spec. RFC 2119 keywords (MUST, SHOULD, etc.) appear sparingly, and mainly in isolated constraints (e.g., "Agent Root MUST NOT store any business agent data") with no corresponding normative framework.

The Terminology section defines five core terms but the definitions are somewhat circular and rely on the architecture not yet introduced. The Problem Statement section (§3) is very brief and mostly restates the abstract. The Security Considerations, IANA Considerations, and Acknowledgements sections are all placeholders.

For a -00 draft this is not unusual, but the architectural sections themselves would benefit from considerably more precision before the DAWN working group could meaningfully evaluate the proposal.

Background of the Authors

Jiankang Yao (CNNIC) — 8 published RFCs and 5 active drafts. An experienced IETF contributor who chairs the Technology Deep Dive (TDD) group and serves as an ART Area Review Team reviewer. Not a newcomer.

Guanggang Geng (Jinan University) — 1 published RFC (RFC 9455, 2023) and 3 active drafts spanning DAWN, ADD, IPWAVE, and DPRISE. Moderately experienced.

Meiling Chen (China Mobile) — 2 published RFCs and 17 active drafts across multiple working groups (ipsecme, spring, lake, lamps, bmwg). Very active IETF contributor, though spread across many unrelated areas.

Hongtao Li (CNNIC) — 1 published RFC (RFC 9095, 2021) and 2 active drafts (regext, dawn). Limited IETF experience.

The author team is primarily from Chinese institutions (CNNIC, Jinan University, China Mobile). Jiankang Yao brings meaningful IETF experience, but the co-authors are generally early-stage contributors. The breadth of Meiling Chen's activity across unrelated working groups may indicate more opportunistic participation than deep domain expertise in agent discovery specifically.