Multi-agent AI communication: new theory reveals tradeoffs between agent count and bandwidth

A new study from Michael Rizvi-Martel and co-authors (Stanford, McGill, and others) tackles a fundamental question in AI: how to best structure multi-agent systems where multiple LLMs collaborate on complex reasoning tasks. Published on arXiv (2510.13903v2), the paper provides a theoretical framework to analyze expressivity—what such systems can and cannot compute—across three algorithmic families: state tracking, recall, and k-hop reasoning. The authors derive precise bounds on three key resources: the minimum number of agents needed to solve a task exactly, the quantity and structure of inter-agent communication required, and the speedup possible as problem size and context length grow.

The framework reveals regimes where communication is provably beneficial and identifies intrinsic limitations when either agent count or bandwidth is constrained. For example, tasks with long-range dependencies demand more communication; increasing agents can reduce bandwidth but only up to a point. The researchers validate their theory through controlled synthetic benchmarks on pretrained LLMs, confirming predicted tradeoffs. The 34-page analysis, with 14 figures, offers principled, actionable guidance for engineers designing scalable multi-agent reasoning systems—a near-term solution to the degradation of single-model performance on long-context, complex problems.