Artificial intelligence is evolving from centralized models running in isolated platforms toward distributed populations of autonomous agents operating across devices, edge infrastructures, cloud systems, robots, vehicles, digital twins, and cyber-physical environments. These agents will exchange not only data, but also models, knowledge, decisions, and actions. They will learn collaboratively, adapt to changing conditions, and coordinate their behaviour, giving rise to what can be described as an Internet of AI Agents.

This keynote examines the architectural foundations required to support distributed learning and inference across heterogeneous computing and communication infrastructures. It will discuss approaches such as federated and collaborative learning, split inference, model and task partitioning, knowledge exchange, agent discovery, semantic interoperability, and the joint orchestration of communication, computation, storage, and energy resources.

The talk will also explore why current solutions may be insufficient for ecosystems composed of autonomous and continuously evolving entities. Such systems introduce demanding requirements in terms of latency, reliability, scalability, energy efficiency, security, trust, provenance, and accountability. In many applications, the network must become more than a transport infrastructure: it must provide an intelligent coordination fabric supporting the interaction and lifecycle of distributed agents.