RecursiveMAS cuts multi-agent AI costs by 75%: researchers

Researchers from UIUC and Stanford introduced RecursiveMAS, a recursive multi-agent AI framework that communicates through latent embeddings instead of text.

In tests, the system improved accuracy across complex tasks, sped up inference, and used fewer tokens than text-based agent chains.

RecursiveMAS also proved cheaper to train than common fine-tuning approaches like LoRA, making custom multi-agent systems more practical to build and scale.