RecursiveMAS speeds up multi-agent AI inference

Researchers have developed RecursiveMAS, a framework that enhances multi-agent AI systems by enabling agents to communicate via embedding space instead of text. This innovation significantly reduces latency, cuts token costs by up to 75%, and improves inference speed by 1.2x to 2.4x. RecursiveMAS allows agents to collaborate and refine reasoning iteratively in the latent space, with only the final agent producing a text output. This approach is more cost-effective for training and requires less GPU memory compared to traditional methods, making complex multi-agent workflows more viable for enterprise deployment.