On the Reliability Limits of LLM-Based Multi-Agent Planning

TL;DR

This paper analyzes the fundamental reliability limits of LLM-based multi-agent planning, modeling it as a decision network and quantifying the impact of communication constraints on decision quality.

Contribution

It provides a theoretical framework linking communication, information compression, and decision quality in LLM-based multi-agent systems, revealing fundamental limits.

Findings

Delegated multi-agent planning is decision-theoretically dominated by centralized decision-making.

Optimal communication strategies can be characterized as budget-constrained experiments on shared signals.

The information loss due to communication can be quantified using divergence measures under proper scoring rules.

Abstract

This technical note studies the reliability limits of LLM-based multi-agent planning as a delegated decision problem. We model the LLM-based multi-agent architecture as a finite acyclic decision network in which multiple stages process shared model-context information, communicate through language interfaces with limited capacity, and may invoke human review. We show that, without new exogenous signals, any delegated network is decision-theoretically dominated by a centralized Bayes decision maker with access to the same information. In the common-evidence regime, this implies that optimizing over multi-agent directed acyclic graphs under a finite communication budget can be recast as choosing a budget-constrained stochastic experiment on the shared signal. We also characterize the loss induced by communication and information compression. Under proper scoring rules, the gap between the centralized Bayes value and the value after communication admits an expected posterior divergence representation, which reduces to conditional mutual information under logarithmic loss and to expected squared posterior error under the Brier score. These results characterize the fundamental reliability limits of delegated LLM planning. Experiments with LLMs on a controlled problem set further demonstrate these characterizations.