A team of researchers published RecursiveMAS, a framework that replaces text-based inter-agent communication with a unified latent-space recursive computation. A lightweight RecursiveLink module links heterogeneous agents into a collaboration loop, enabling shared internal state propagation and end-to-end gradient-based co-optimization across recursion rounds. Evaluated on 9 benchmarks spanning mathematics, science, medicine, search, and code generation, RecursiveMAS achieves an average 8.3% accuracy gain over strong baselines, 1.2–2.4× inference speedup, and a 34.6–75.6% reduction in token usage—suggesting that structured latent-space communication is a more efficient substrate for multi-agent collaboration than serialized language exchange.