Charter for Research Group

Internetworking Challenges for AI (Net4AI)

Motivation

In the past few months, the emergence of generative AI has triggered a lot of interest and activities in the IETF community, as shown by the number of BoF organized on the topic, giving raise to BoFs such as Discovery of Agents, Workloads, and Named entities (DAWN), Dynamic Multi-agent Secured Collaboration (DMSC) or Agent Communication Protocols (AGENTPROTO) and increased activity in established working groups such as OAUTH or WIMSE to address specific requirements of generative or agentic AI systems.

While those initiatives target immediate or mid-term concerns related to distributed generative and agentic AI systems, there is a need for a principled, clean slate discussion about what would be needed for distributed and multi-stakeholder AI over the Internet. Net4AI's goal with regards to the Internet community is to be the research group hosting and facilitating this discussion.

In the research world, the interest for networking aspects of AI system is growing in major conference venues. Major networking and distributed systems conferences now feature dedicated tracks focusing on training and inference networking, LLM serving systems or LLMs used as networking toolsshoing a growing interest for distributed AI systems.

Yet, some topics are in a blind spot for now, such as the architectural Internet support for autonomous agents, principled decentralization, network architectures for agents, Internet-native support for AI ecosystems. Net4AI's goal with regards to the research community is to gather a community to work on this blind spot and trigger more research in collaboration with the Interent community.

Topics

Possible research topics for The Net4AI research group include, but are not limited to:
- Usage and adaptation of internet protocols to operate AI workloads;
- Testbeds and benchmarking to foster idea exchange between research projects on agentic management of networks;
- Agent-specific communication protocols beyond real time human-readable communications;
- Caching / content distribution for AI workloads, towards KV-centric networking at scale to distribute inference workloads beyond a datacenter’s realm;
- Principled agent identity, naming and addressing, encompassing cross
domain and decentralized identity, as a foundation for trustworthy
agent to agent communications;
- New security requirements of Agentic AI.

Way of Working

Work in the Net4AI research group will be facilitated by several activities:
- Assessing & developing experimental tools (simulators, testbeds)
- Collecting and sharing datasets for reproducible experiments
- Driving research on networked systems for AI (not limited to datacenters)
- New architectures (KV-Cache-Centric, Mixture-of-Experts)
- New requirements (inter-DC communication, security)
- Developing principled approaches on Agentic Communication
(e.g., “Internet of Agents”)