Control-theoretic governance layer improves multi-LLM agent collaboration by 32 percentage points in simulated financial services environment

Researchers propose a control-theoretic governance layer called the Experience Orchestrator (EO) to address the collapse of multi-LLM agent systems when agents with structurally opposed objectives interact over multiple turns. In a simulated financial services environment, EO guides a site agent and a visitor agent—designed with opposing goals—toward a shared outcome: high-intent advisor contact.

EO integrates three mechanisms: a Contextual Bandit (CB) that selects content arms using real-world web analytics, a PID controller that enforces behavioral consistency via dynamic schema constraints, and a POMDP belief tracker that maintains a probabilistic model of visitor intent. Across 60,000 simulations, EO achieved a 32 percentage point improvement in high-intent advisor contact rates compared to a naive LLM control (78.1% vs. 46.1%).

The authors report that the CB’s variant selection accounted for 97% of the between-factor outcome variance, indicating that the governance policy—not initial environmental conditions—determined the trajectory outcomes. Persona-level analysis revealed two regimes: for visitors with no natural inclination toward conversion, the governance layer was critical to system functionality, while visitors already near alignment saw limited benefit from EO.

All findings are conditional on LLM-to-LLM simulation, and the PID controller has not been calibrated against real human unpredictability. The authors emphasize that validating EO on live traffic is the critical next step.