Princeton's Kilian Lieret highlights ProgramBench data showing multi-agent systems fail to outperform single agents but reach mediocre solutions twice as fast

This trade-off prioritizes generation speed over final code optimization.