When an agent becomes more than an agent

The article discusses growing concerns among researchers regarding the unpredictable behavior of LLM-based agents, which could pose threats to both digital platforms and society. The author analyzes calls to halt model development until strict safety protocols are implemented, expressing skepticism about the effectiveness of such measures. Drawing parallels between anthills, human communities, and AI systems, the author argues that dangerous properties emerge not from the agent itself, but from the complex interaction of the 'agent plus environment' system. According to the author, attempts to control behavior at the individual agent level are destined to fail, as the overall direction of the system is not determined by the agent alone. The article calls for a rethinking of AI safety approaches, shifting the focus from individual agent characteristics to systemic emergent properties that cannot be fully predicted or constrained by current development methods.