Scaling Large-Language-Model-based Multi-Agent Collaboration