Control-Theoretic Governance for Multi-LLM Agent Conversations

An arXiv paper (2608.11207) introduces a governance framework grounded in control theory to oversee interactions among various large language model (LLM) agents with conflicting aims. Conducted in a simulated financial services context, the research tackles the issue of conversational collapse that arises when two agents with fundamentally opposing objectives engage without a common goal function. The proposed solution, termed the Experience Orchestrator (EO), integrates three components: a Contextual Bandit (CB) for content selection based on real-world web analytics, a PID controller to ensure behavioral consistency through dynamic schema constraints, and a POMDP belief tracker for sustaining a probabilistic model of visitor states. The findings indicate that this governance layer can effectively replace the absent goal function, averting collapse and fulfilling specified objectives. The paper can be accessed on arXiv with the identifier 2608.11207.