Silo-Bench: 'Communication-Reasoning Gap' — Scaling Agent Count Does Not Bypass Individual Context Limits, Coordination Overhead Eventually Dominates

arXiv·medium signalXBlueskyLinkedInCopy link

Across 1,620 experiments over 54 configurations and 30 algorithmic tasks, Zhang et al. find that LLM agents successfully form coordination topologies and share information, but consistently fail when integrating distributed information into correct answers. The 'Communication-Reasoning Gap' means more agents cannot compensate for individual context limit failures — coordination overhead eventually eliminates parallelization gains. Critical counterevidence for the common assumption that multi-agent architectures solve context-length problems.