The big picture

Multi-agent systems still lack reliable coordination at scale. Open-ended social tasks expose failures that controlled benchmarks routinely hide.