Researchers at Princeton University have shown that a multi-agent approach achieves approximately 70% accuracy on SWE-bench, a demanding benchmark for software engineering and reasoning tasks. This method, called a “marketplace of agents,” relies on combining specialized agents developed independently by different creators, rather than on a single general-purpose model. The results suggest that specialization combined with composition outperforms generalization at scale. This finding could influence the competition between OpenAI, Anthropic, and Google DeepMind, which are currently vying for AI agent performance leadership. The authors acknowledge, however, that benchmarks remain controlled environments and that coordinating multiple agents adds potential failure points.
Source: Read the original article
