AgentConductor's AI team beats benchmarks with 14.6% better code generation

A research team led by Siyu Wang developed AgentConductor, a multi-agent system (MAS) that uses a reinforcement learning-optimized orchestrator to dynamically generate interaction topologies. It creates task-adapted, layered DAGs based on inferred difficulty, reducing redundant communication. On competition-level code datasets, it achieved state-of-the-art pass@1 accuracy, outperforming the strongest baseline by up to 14.6% while significantly cutting token costs and communication density.