Patterns and problems in emerging multi-agent systems

As AI agents become increasingly capable and prevalent in shared systems like codebases and markets, real-world agent-agent interactions are expected to scale rapidly, potentially exceeding human interactions before institutions fully understand how to manage them safely. Current AI agents excel at tool use and individual tasks but struggle with genuine peer-to-peer coordination and can exhibit problematic behaviors like confabulation and reward hacking that may compound into systemic failures. The research identifies these behavioral risks in frontier models and calls for developing better understanding and mitigation strategies before multi-agent systems become widespread.