Dynamic Governance Boosts Multi-LLM Agent Conversational Outcomes.

When multiple LLM agents with conflicting objectives interact, conversations often collapse without achieving either agent's goals. This research explores whether a control-theoretic governance layer can resolve this by substituting a missing shared goal function. The proposed Experience Orchestrator (EO) framework was tested in a simulated financial services scenario where a site agent aimed to guide a visitor to an advisor, while the visitor maintained realistic resistance. EO employs three mechanisms: a Contextual Bandit (CB) for content selection, a PID controller for behavioral consistency, and a POMDP belief tracker for visitor intent. Across 60,000 simulations, EO dramatically improved the high-intent advisor contact rate by 32 percentage points compared to a naive LLM control. The Contextual Bandit variant selection accounted for 97% of the outcome variance, confirming the governance policy's critical role. The study also found that governance is essential for visitors with low natural inclination to convert, while empathetic LLM defaults suffice for aligned visitors. The findings are conditional on LLM-to-LLM simulation, with real-world human validation as the next step.