The development of agentic artificial intelligence — systems capable of planning, using tools and acting autonomously in complex environments — is advancing faster than the theoretical understanding of how these systems interact with their surroundings and with other agents. A comment published in Nature argues that a dedicated discipline is needed to study those interactions, rather than continuing to extrapolate metrics designed for isolated language models.

The central claim is that an agent's behaviour cannot be understood from its internal capabilities alone: it depends critically on how the agent negotiates, cooperates or competes with other agents, humans and systems. Without a formal framework to describe these dynamics, it becomes difficult to predict failures, design reliable evaluations or establish safety guarantees when multiple agents operate simultaneously in the same environment.

The authors propose that this "science of interaction" combine tools from game theory, multi-agent reinforcement learning, complex systems theory and cognitive science. The goal would be to develop models describing how patterns of cooperation, conflict and coordination emerge, and how small changes in a system's configuration can trigger behaviours not anticipated by its designers.

The proposal has immediate practical implications. As AI agents are integrated into critical infrastructure, financial markets, scientific research and everyday devices, the ability to anticipate and control their interactions becomes a safety requirement rather than an academic luxury. The authors call on the research community to prioritise this approach before the scale of deployment outstrips the capacity for human oversight.