Two LLM agents with structurally opposed goals, left to interact over multiple turns, do not produce productive competition: they collapse. The visitor gives in, the site agent stops varying its approach, and the conversation ends without either side reaching its stated objective. That failure is the starting point for Experience Orchestrator (EO), a governance layer based on control theory, tested in a simulated financial services environment.

EO does not add a larger model or a more persuasive prompt. It inserts explicit control over the joint trajectory: a Contextual Bandit selects content variants calibrated from real-world web analytics; a PID controller enforces behavioral consistency through dynamic schema constraints; a POMDP belief tracker maintains a probabilistic model of visitor intent.

Across 60,000 simulations, the high-intent advisor contact rate rises from 46.1% in the naive LLM control to 78.1% with EO: a 32 percentage point lift. The most relevant finding is not the absolute result. Variance analysis shows that Contextual Bandit variant selection accounts for 97% of the between-factor variance. In other words, it is the governance policy, not the initial state of the environment, that determines where trajectories end up.

That is the structural point. In the current wave of LLM agents, cooperation is often taken for granted: the assumption is that sufficiently capable models will find a communicative equilibrium. The experiment shows the opposite when objectives are opposed. Without a shared goal function, the emergent behavior is not competition but surrender. Governance is not decoration: it is the functional substitute for that missing goal function.

Persona-level analysis reveals two distinct regimes. For visitors with no natural inclination toward conversion, the governance layer makes the difference between a functional and a non-functional system. For visitors already near alignment, a naive LLM's empathetic defaults are largely sufficient. This has second-order implications: not every conversational flow requires the same intensity of control. Companies designing financial services assistants may need to reserve active control for high-friction segments and leave simpler cases to lighter interaction.

The work is conditional on LLM-to-LLM simulation. The PID controller has not been calibrated against real human unpredictability; validating EO on live traffic is indicated as the critical next step. That constraint weighs doubly in financial contexts: real user behavior changes over time and calibration data must stay current. Organizations running these systems self-hosted can keep the governance loop and behavioral data within their own boundaries, but must accept that the advantage observed in simulations does not automatically transfer to production.

From a deployment standpoint, EO introduces no specific hardware requirements: its value lies in controlling interaction, not in expanding inference capacity. For an organization already running LLMs on-premise, a layer of this kind sits upstream or alongside orchestration, adding governance surface rather than computational load. The question then becomes one of auditability: a POMDP belief tracker produces probabilistic states that can be logged and verified, a relevant aspect in conversations that push a user toward an advisor.

It remains to be seen whether PID- and POMDP-based control holds up against human variability. The paper identifies this as the critical next step, and not by chance: 60,000 simulations can reveal much about agent-to-agent dynamics, but almost nothing about the moment a real user decides to stop responding.