AI Agents Talk Without Talking: Faster, Cheaper

Researchers have introduced RecursiveMAS, a groundbreaking framework that allows AI agents to collaborate by transmitting information through embedding space, bypassing the inefficiencies of text-based communication. This approach significantly reduces latency and token costs, thereby enhancing both the speed and performance of multi-agent systems.

Unlike traditional methods that rely on sequential text generation, RecursiveMAS enables agents to pass continuous latent representations. This method allows the entire system to refine collective reasoning over multiple rounds entirely in the latent space, with only the final agent producing a textual output. This simulates telepathic communication among agents.

The framework utilizes a specialized component called RecursiveLink, which transmits and refines latent states without decoding text. This innovative design keeps the core language models frozen, optimizing the system by training only the lightweight RecursiveLink parameters. This architecture also allows multiple agents to share a single backbone model, reducing memory requirements.

Evaluations across nine benchmarks showed RecursiveMAS achieved an average accuracy improvement of 8.3% and up to 2.4x faster inference speeds. It also reduced token usage by up to 75.6% and cut training costs by more than half compared to full fine-tuning, making complex multi-step agent workflows more viable for enterprise deployment.