Summary

  • The IRTF posted the first Net4AI charter on 2 October and placed it in initial internal Steering Group/IAB review. Datatracker lists the group as Proposed, not as a chartered research group; no milestones are listed.
  • The draft would convene research on distributed AI over the Internet, including workload networking, reproducible testbeds, agent communication and identity, and KV-centric distribution. It is a research proposal, not an approved standards program.

The immediate AI agenda in Internet standards circles tends to appear as a collection of protocol questions: how agents discover one another, exchange messages or apply existing security machinery. Net4AI’s first proposed charter asks for a wider frame. Its authors want a research venue to consider how distributed, multi-stakeholder AI might fit into the Internet’s architecture at all.

Datatracker records charter-irtf-net4ai-00-00 as a new document on 2 October 2026. The proposal is at “Start Chartering/Rechartering (Internal Steering Group/IAB Review),” and the group’s state is “Proposed.” There is no active ballot, telechat date or milestone schedule. This is a formation proposal—not evidence that the IRTF has approved a new group or that the named work is underway.

The distinction matters because the charter is deliberately broad. It names adapting Internet protocols for AI workloads, testbeds and benchmarking, agent-specific communication beyond human-readable real-time exchanges, and caching or content distribution for AI workloads. One example is KV-centric networking intended to move inference workloads beyond a single data centre. It also identifies decentralized and cross-domain agent identity, naming and addressing, alongside new security requirements for agentic AI.

The work plan is equally exploratory: simulators and testbeds, reproducible datasets, networked-systems research beyond data centres, and architectures such as KV-cache-centric and Mixture-of-Experts designs. The charter lists no milestones. Readers therefore cannot infer a delivery date, protocol commitment or implementation result from the breadth of the agenda.

Net4AI’s motivation acknowledges existing IETF activity, including DAWN, DMSC, AGENTPROTO, OAUTH and WIMSE. The proposed research group is not presented as replacing those efforts. Instead, the charter argues that near-term protocol work leaves a longer-horizon question open: what network architecture should support AI systems distributed across organizations and the wider Internet?

That is a different institutional job from writing a standards-track protocol. RFC 2014 describes IRTF groups as places to investigate longer-term Internet research; RFC 7418 explains that their charters sketch a research field and that their outputs need not be RFCs. The current Net4AI proposal has not cleared the first review gate, but it makes the boundary visible: a research agenda can explore workload placement, trust and system architecture without promising that any one design will become an Internet standard.

For network operators and researchers, the practical signal is not a new deployment instruction. It is a proposed venue whose scope reaches from datacentre-scale inference distribution to decentralized agent naming. Whether that breadth becomes a useful research program will depend first on charter approval and then on how participants turn it into questions, evidence and a workable agenda.

Sources