Dynamic Governance of Multi-LLM Agent Systems for Collaborative Conversational Outcomes

Abstract

When two large language model (LLM) agents with structurally opposed objectives interact over multiple turns, the absence of a shared goal function leads not to competition but to collapse: the visiting agent capitulates, the site agent ceases to vary its approach, and the conversation terminates without achieving either party's stated objective. This paper investigates whether a control-theoretic governance layer can substitute for this missing goal function.

The Experience Orchestrator (EO) addresses this challenge in a simulated financial services environment, where a site agent guides a visitor toward advisor contact while the visitor exhibits psychologically realistic resistance. EO governs the joint conversation trajectory through three integrated mechanisms: a Contextual Bandit (CB) that selects content arms calibrated from real-world web analytics data, 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 improvement in high-intent advisor contact rate (78.1% versus 46.1% over a naive LLM control). Notably, CB variant selection accounts for 97% of between-factor outcome variance, confirming that the governance policy—rather than environmental initial conditions—determines where trajectories ultimately converge.

Persona-level analysis reveals two distinct operational regimes. For visitors with no natural inclination toward conversion, the governance layer is the decisive factor between a functional and non-functional system. Conversely, for visitors already near alignment, a naive LLM's empathetic defaults prove largely sufficient.

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

Overview

This paper, submitted to the AI Engineer World's Fair 2026, is authored by Alexander Liss, Nicholas Desmond, and Santiago Gil Gallego. It comprises 13 pages, 3 figures, and 3 tables, and is available under arXiv:2608.11207 in the Artificial Intelligence (cs.AI) category.

Key Contributions

A novel governance framework: The Experience Orchestrator integrates control-theoretic methods (CB, PID, POMDP) to manage multi-turn LLM agent conversations where objectives conflict.
Empirical validation: A large-scale simulation study (60,000 runs) demonstrates substantial improvements in outcome achievement over naive LLM baselines.
Factor analysis: Identification of CB variant selection as the dominant driver of outcome variance, offering actionable insights for system design.
Persona-specific findings: Distinction between visitor profiles where governance is essential versus where simpler approaches suffice, guiding deployment strategies.

Methods

The EO framework operates through three layers:

Contextual Bandit (CB): Selects content strategies (arms) based on contextual features, calibrated using real-world web analytics to optimize engagement.
PID Controller: Applies proportional-integral-derivative control to maintain behavioral consistency, dynamically adjusting schema constraints to keep the conversation on track.
POMDP Belief Tracker: Maintains a probabilistic belief state over visitor intent, enabling the system to adapt its approach under uncertainty.

Results

High-intent conversion rate: EO achieves 78.1% versus 46.1% for the naive LLM control, a +32 percentage point improvement.
Variance attribution: CB variant selection explains 97% of between-factor outcome variance, highlighting the importance of content selection over environmental factors.
Regime analysis: Two distinct visitor profiles emerge—those requiring robust governance for any conversion (low-alignment) and those for whom empathetic defaults suffice (high-alignment).

Limitations and Future Work

Simulation dependency: All results are based on LLM-to-LLM simulations, which may not fully capture human conversational unpredictability.
Uncalibrated PID: The PID controller requires validation against real user behavior.
Next steps: Live-traffic validation of the EO system is essential to confirm its effectiveness in real-world settings.

Conclusion

The Experience Orchestrator demonstrates that a control-theoretic governance layer can effectively substitute for a missing shared goal function in multi-LLM agent systems, significantly improving collaborative outcomes. These findings provide a foundation for designing more robust conversational AI systems, with clear directives for future deployment and calibration.

via ArXiv AI