The Information-Theoretic Cost of Hierarchy: A Channel-Theoretic Analysis of LLM Multi-Agent Organizations

Read full text Summarise with VIRAT
Authors: Koh Shinoda
Cited by: 0

Summary

No summary yet. VIRAT can read this paper and write a plain-language summary of its aim, methods, key findings and limitations.

Summarise with VIRAT Free to try with your Science Catalogs account.

Abstract

This paper investigates the impact of organizational structure in LLM-based multi-agent systems from an information-theoretic perspective. We model agents as communication channels and analyze information flow using synthetic tracer injection. Our results reveal that intermediate management agents (Team Leads) destroy 71–97% of task-relevant information in a single relay step, acting as information bottlenecks rather than coordinators. Through 186 controlled experiments, we demonstrate that shallow organizations outperform deep hierarchies due to reduced information distortion. We further decompose information loss into filtering and channel noise, and show that high-capacity models mitigate depth-related degradation. Importantly, we establish that high-performing systems do not preserve information, but transform it through augmentation. These findings suggest that LLM multi-agent systems should be designed as information transformation networks rather than information preservation pipelines.

Fields of Study & Subjects

Multi-Agent Systems and Negotiation Complex Systems and Dynamics Diverse Interdisciplinary Research Studies

Publication details

Journal / source
Open Science Framework
Publication date
1 January 2026
Language
EN