Experience Orchestrator boosts LLM agent conversions 32 points in 60k simulations

A new paper from Alexander Liss, Nicholas Desmond, and Santiago Gil Gallego (arXiv:2608.11207) tackles a core problem in multi-agent LLM systems: when two agents with opposing objectives interact over multiple turns, the conversation often collapses instead of competing. The authors propose the Experience Orchestrator (EO), a control-theoretic governance layer tested in a simulated financial services environment where a site agent tries to guide a visitor toward advisor contact, while the visitor exhibits realistic resistance. EO operates via three mechanisms: a Contextual Bandit that selects content arms calibrated from real web analytics, a PID controller that enforces behavioral consistency through dynamic schema constraints, and a POMDP belief tracker that maintains a probabilistic model of visitor intent. The governance layer aims to replace the missing shared goal function that is absent in open-ended LLM dialogues.

Across 60,000 simulations, EO delivers a +32 percentage point lift in high-intent advisor contact rate—78.1% vs. 46.1% for a naive LLM control. Notably, Contextual Bandit variant selection accounts for 97% of between-factor outcome variance, suggesting the governance policy, not initial conditions, determines where trajectories end. Persona-level analysis reveals two regimes: visitors with no natural inclination to convert need governance to function at all, while near-aligned visitors already respond well to empathetic defaults. The authors caution that all findings are based on LLM-to-LLM simulation; the PID controller has not been calibrated against human unpredictability, making validation on live traffic the critical next step.