Dynamic Governance of Multi-LLM Agent Systems for Collaborative Conversational Outcomes
06:00 · August 13, 2026 · arXiv cs.AI RSS
When two LLM agents with structurally opposed objectives interact across multiple turns, the absence of a shared goal function produces not competition but collapse: the visitor capitulates, the site agent stops varying its approach, and the conversation terminates without achieving either agent's stated objective. This paper asks whether a control-theoretic governance layer can substitute for that missing goal function. The Experience Orchestrator (EO) addresses this in a simulated financial services environment where a site agent guides a visitor toward advisor contact while the visitor maintains psychologically realistic resistance. EO governs the joint trajectory through three mechanisms: a Contextual Bandit (CB) that selects content arms calibrated from 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 achieves a +32 percentage point lift in high-intent advisor contact rate (78.1% vs. 46.1% over a naive LLM control), with CB variant selection accounting for 97% of between-factor outcome variance -- confirming that the governance policy, not environmental initial conditions, determines where trajectories end up. Persona-level analysis reveals two distinct regimes: for visitors with no natural inclination toward conversion, the governance layer is the difference between a functional system and a non-functional one; for visitors already near alignment, a naive LLM's empathetic defaults are largely sufficient. All findings are conditional on LLM-to-LLM simulation. The PID controller has not been calibrated against real human unpredictability, and validating EO on live traffic is the critical next step.
Summary
When two LLM agents pursue structurally opposed objectives across multiple conversational turns, the lack of a shared optimization target often leads to collapse rather than productive negotiation. Each agent converges on agreement regardless of its initial goals, producing terminal states that satisfy neither party. The Experience Orchestrator (EO) addresses this architectural gap by inserting an external governance layer grounded in control theory, rather than attempting to engineer a joint reward function inside the models themselves.
EO operates in a simulated financial-services setting in which a site agent seeks to move a visitor toward scheduling an advisor consultation while the visitor agent maintains persona-driven resistance. Governance is achieved through three coordinated mechanisms. A contextual bandit selects content variants whose probabilities are derived from real-world web analytics. A PID controller applies dynamic schema constraints to maintain behavioral consistency across turns. A POMDP belief tracker maintains an explicit probability distribution over the visitor’s latent intent states, enabling the system to update its model of the visitor at every step.
In a factorial evaluation comprising 60,000 LLM-to-LLM simulations, the full EO configuration produced a 32-percentage-point increase in high-intent advisor contact rate relative to a naive LLM baseline governed only by system prompts. Variant selection performed by the contextual bandit accounted for 97 percent of outcome variance, indicating that the governance policy, rather than initial environmental conditions, primarily determines trajectory success. Persona-level breakdowns further show that the governance layer is decisive for visitors lacking natural conversion inclination, while visitors already near alignment require little additional steering.
The reported gains remain conditional on simulation. The PID controller has not been tuned against the higher variance of actual human behavior, and the authors identify live-traffic validation as the necessary next step before deployment.
Why it matters
This research is highly relevant for Dutch AI researchers and enterprise practitioners, particularly in the financial and customer service sectors, as it offers a novel, mathematically grounded framework for governing autonomous LLM agents. Its focus on external control mechanisms aligns well with EU regulatory demands for predictable and transparent AI behavior.
