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

Review · August 21, 2026

Zero-Configuration Agent Discovery via mDNS

draft-jakab-dawn-agent-discovery-mdns -00View on Datatracker →

Problem Statement

Current agent-to-agent communication protocols (A2A, MCP) assume the connecting party already knows the agent's hostname or URL. This spec addresses the gap for local-network environments — developer workstations, branch offices, meeting rooms, edge deployments — where agents start, stop, and move dynamically and where there is no central registry or manual configuration available. The proposed solution is zero-configuration discovery using existing mDNS (RFC 6762) and DNS-SD (RFC 6763) mechanisms.

How the Spec Fits In

This is a DAWN working group draft, not an OAuth draft. It operates entirely at the network discovery layer and has no dependency on OAuth or authorization specifications. It builds purely on existing DNS infrastructure standards (RFC 6762, 6763, 6335) and uses the Agent2Agent (A2A) protocol as a worked example. The spec adds no new protocol machinery — it profiles existing mDNS/DNS-SD conventions for the AI agent discovery use case and defines a shared service type and two-stage discovery pattern. It is a relatively lightweight document that extends well-understood technology into a new domain.

High-Level Summary

The spec defines a pattern for AI agents to advertise themselves on a local network and for other agents or clients to find them, using the same multicast DNS technology already used for printers, AirPlay speakers, and similar zero-configuration devices.

The core design has three elements:

  • A single shared service type _agent._tcp for all agents regardless of their protocol. The specific agent protocol (e.g., A2A, MCP) is carried in a proto key inside the DNS TXT record. This means a discovering party can browse one service type and then filter by protocol.

  • A minimal DNS record set per agent: a service instance name, an SRV record (host + port), and a TXT record with a small set of key/value pairs. The TXT record carries just enough to identify the agent and point to a full description document.

  • A two-stage discovery flow: Stage 1 uses DNS browsing to enumerate candidates and apply lightweight filtering without any transport-level contact. Stage 2 fetches the full agent metadata document (e.g., an A2A AgentCard) over HTTP/HTTPS and then initiates a session using the agent's native protocol.

Security is explicitly out of scope for the DNS layer. mDNS advertisements are unauthenticated, and the spec correctly notes that trust is established in Stage 2 via TLS and the agent protocol's own authentication mechanisms. TXT fields like caps (capabilities) and agentId are advisory only and cannot be trusted without Stage 2 verification.

The A2A instantiation section provides a concrete field mapping from AgentCard properties to DNS-SD records and examples of what the DNS records look like in practice.

Interoperability

Significant gap: no normative language in requirements. Section 4 defines twelve requirements but explicitly describes them as "stated descriptively rather than in MUST/SHOULD/MAY terms." This is unusual for an IETF specification and weakens the document's ability to produce interoperable implementations. Requirements like REQ-6 (provide enough info to initiate contact) and REQ-8 (avoid exposing sensitive details) have real interoperability implications that should be expressed normatively.

No proto registry yet. The IANA Considerations section acknowledges that a registry for proto key values should be created if the document advances. Without this registry, different agent protocols could independently choose conflicting proto values, undermining the shared-type model. This should be addressed early.

Design tension with LAD-A2A. The Related Work section candidly notes that the LAD-A2A project uses an A2A-specific _a2a._tcp service type rather than the generic _agent._tcp defined here. This is a significant design fork: the shared-type-plus-proto-discriminator model in this spec requires discovering parties to browse a single type and then filter, while the LAD-A2A model allows protocol-specific browsing. The spec argues for the shared-type approach but does not resolve which will prevail in practice — two competing approaches in deployment weakens interoperability.

Mechanisms that could be removed or simplified. The SVCB alternative (Section 5.2) is presented as a forward-looking option but explicitly noted as not yet widely supported in mDNS implementations. Since TXT records are the required fallback, the SVCB discussion adds complexity without adding near-term interoperability value and could be moved to a separate or future document.

Writing Quality

The document is generally well-written and clearly organized for an early individual draft. The abstract and introduction are concise, the two-stage model is explained clearly, and the A2A example records in Section 6.1.6 are helpful. The authors are appropriately candid about competing approaches in the Related Work section.

The use of tables to map AgentCard fields to DNS-SD records is effective and readable.

The main quality gap is the deliberate avoidance of normative language in the requirements section. While the authors may intend to add BCP 14 language in a later revision, the current draft reads more like an informational use-case document than a protocol specification in the critical requirements and solution sections. For a working group adoption discussion, this will likely be flagged.

The IANA Considerations section is appropriately forward-looking about what registrations would be needed but leaves the actual registration requests for later, which is acceptable at -00.

Background of the Authors

Loránd Jakab (Cisco, [email protected]) has 1 published RFC (RFC 7215, LISP deployment considerations, 2014) and 1 active draft — this one. His background is in LISP routing architecture. He is an experienced IETF participant but not a prolific multi-area author. His move into AI agent discovery represents a domain shift from his prior routing work.

Frank Brockners (Cisco, [email protected]) has 10 published RFCs, primarily in the IPPM working group covering In Situ OAM (IOAM). His most-cited work is RFC 9197 (IOAM data fields). He has 6 active drafts and 41 expired drafts, indicating sustained and broad IETF engagement. He is not new to the IETF and brings significant standards process experience, though his background is in network measurement/telemetry rather than application-layer discovery or AI agent protocols.

Both authors are from Cisco, and all four contributors listed are also from Cisco. The document reflects a single-organization perspective, which is worth noting for working group adoption discussions.