Agentic AI needs a science of interaction

Artificial intelligence is moving from stand-alone models towards systems of interacting agents. These systems do not merely classify, predict or generate: they plan, exchange messages, call tools, divide labour, adapt and coordinate action across tasks and time. This shift has made agentic AI one of the dominant frames for the next phase of the field. Understanding how these agents interact will therefore be as important as improving what each agent can do alone. Here, the authors argue that agentic AI requires a shift from evaluating individual agents alone to analyzing interaction architectures, proposing a framework that links agent interactions, collective configurations and time-dependent outcomes.