IETF 126: IPv6 and DNS as AI Agent Foundations

At IETF 126, discussions across working groups pointed to IPv6 and DNS as the essential foundations for a future internet filled with AI agents. The meeting, held by the Internet Engineering Task Force (IETF), the open standards body where engineers agree on technical rules for the internet, drew 1,230 on-site participants and 672 remote attendees, setting a new attendance record. Anlei Hu, Chief Scientist of the China Internet Network Information Center (CNNIC) and an APNIC Executive Council member, attended sessions on IP addressing, DNS operations, and AI. In a report for APNIC, Hu wrote that consensus is gradually emerging across the IPv6 Maintenance (6MAN), Distributed Mobility Management (DMM), and DNS Operations (DNSOP) working groups and the Internet Architecture Board (IAB) Open Meeting: IPv6 and DNS will play increasingly important roles in the AI era. IPv6, the newer internet addressing system with a vastly larger pool of unique addresses, provides the scale needed for large AI deployments, while DNS, the internet's address book that turns names into numeric addresses, remains essential for service discovery and connectivity.

On IP addressing, China Telecom presented a proposal called DNS-Based Address Mapping Record (AMR) for IPv4/IPv6 Mapping in IPv6-only Networks. The idea is to use DNS records to carry static IPv4-to-IPv6 mapping information within networks that run only IPv6, the next-generation addressing protocol. This approach aims to improve operations of DNS64 and NAT64, the translation mechanisms that let IPv6-only networks reach IPv4 resources, and to reduce the need for hard-coded mappings. At the IAB Open Meeting, the long-term implications of IPv4 exhaustion were a central topic. Participants broadly agreed that large-scale deployment of AI agents and distributed compute nodes will require traceable, globally unique addresses. The anticipated scale of those deployments exceeds the practical limits of IPv4, the older addressing system with a limited number of addresses that is now largely used up, making IPv6 a fundamental requirement. Several participants also noted that large-scale Network Address Translation (NAT), which lets many devices share one public IP address, and other translation mechanisms disrupt end-to-end identity, complicate agent discovery and trusted communications, and create challenges for security auditing. Separate work continued on Segment Routing over IPv6 (SRv6) and Compressed SRv6, documented in RFC 9800, for compute interconnect environments. Key topics included low-latency IPv6 forwarding and In-situ Operations, Administration, and Maintenance (IOAM) telemetry for GPU clusters and distributed AI training systems.

Most DNS-related discussions at IETF 126 took place within the DNS Operations Working Group (DNSOP), which focuses on how DNS systems are run and improved. Debate continued around proposals for DNS-based AI agent discovery. At the previous IETF 125 meeting, a proposal suggested using the _agents subdomain with Service Binding (SVCB) and TXT records to support intra-domain agent discovery. SVCB records can point a domain name to a service and its connection details, while TXT records are flexible text fields often used for various metadata. Supporters argued that existing mechanisms such as DNS Service Discovery (DNS-SD) and SVCB could support lightweight agent discovery without creating new protocols. However, many DNSOP participants suggested that DNS is not designed to carry complex descriptions of AI agent capabilities. Concerns included the lack of standardized semantic models to describe what an AI agent can do, the inappropriate use of TXT records for structured data, and the potential for increased query volumes that could burden DNS infrastructure. Based on the DNSOP discussions Hu attended, the prevailing view was that DNS should not be used to transport complex capability metadata. Instead, DNS should provide only basic service endpoint information, such as where a service can be reached. More advanced agent discovery mechanisms are expected to develop through the Discovery of Agents, Workloads, and Named entities (DAWN) initiative.

AI and AI agents were prominent topics at IETF 126, and multiple workstreams focused on developing an interoperable internet protocol framework for agent-to-agent communication. The most significant development was the launch of the DAWN Birds of a Feather (BoF) session, an informal meeting where interested engineers gauge support for forming a new working group. DAWN is expected to progress towards formal Working Group status. The initiative aims to create a decentralized, scalable, cross-domain framework for AI agent discovery. Its approach can use DNS as an underlying transport mechanism while avoiding the need to embed complex semantic information within DNS itself. This design addresses many of the concerns raised during the DNSOP discussions. The DAWN session concluded with sufficient support to continue the Working Group formation process, and a formal proposal is expected to be submitted at IETF 127.

Hu's main takeaways were that the internet's existing architecture may need to evolve to support large-scale deployment of AI agents. Many participants suggested that AI could strengthen the case for IPv6, because IPv4 scarcity and increasingly complex NAT environments are poorly suited to trusted, end-to-end communication between agents. There was also broad agreement that DNS should remain focused on naming and service discovery, rather than becoming a repository for detailed AI capability information. Work on agent discovery is likely to progress through the proposed DAWN Working Group, while researchers and operators across the IETF are beginning to examine how AI-generated traffic could change DNS usage patterns and affect infrastructure stability. Hu's interpretation of discussions across several working groups and meetings is that a four-layer model for an AI-enabled internet could emerge: IPv6 for addressing and connectivity, DNS for naming and service discovery, DAWN for agent and workload discovery, and Agent Communication Protocols (agentproto) for agent communication and interoperability. This remains an early-stage vision rather than an agreed architecture, but it illustrates how participants are beginning to think about the protocol foundations needed to support an internet increasingly populated by AI agents.

For the APNIC community, these discussions offer an early view of issues that could become increasingly important. The relationship between AI agents and IP addressing is only beginning to be explored, but the implications are potentially significant. Issues such as identity, traceability, addressing, and security align closely with APNIC's role in supporting internet number resource management, routing security, and the ongoing development of internet infrastructure. Likewise, the potential impact of machine-generated traffic on DNS infra