“Preliminary experimental results suggest that splitting a fixed training compute budget among multiple agents can yield gains compared to allocating it all to a single agent.”