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

多大模型智能体系统的动态治理：实现协作式对话成果

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. 当两个目标结构对立的大模型（LLM）智能体进行多轮交互时，由于缺乏共同的目标函数，结果往往不是竞争，而是系统崩溃：访客选择妥协，站点智能体停止调整策略，最终对话在双方都未达成既定目标的情况下终止。

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. 本文探讨了是否可以通过一个基于控制理论的治理层来替代缺失的目标函数。研究提出了“体验编排器”（Experience Orchestrator, EO），并在一个模拟的金融服务环境中进行了验证。在该环境中，站点智能体引导访客联系顾问，而访客则保持心理学上真实的抵触情绪。

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. EO 通过三种机制来管理交互轨迹：一是利用从真实网络分析中校准的内容臂进行选择的上下文多臂老虎机（CB）；二是通过动态模式约束强制执行行为一致性的 PID 控制器；三是维护访客意图概率模型的 POMDP（部分可观测马尔可夫决策过程）信念追踪器。

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. 在 60,000 次模拟中，EO 将高意向顾问联系率提升了 32 个百分点（达到 78.1%，而原始 LLM 对照组仅为 46.1%）。其中，CB 变体选择占因素间结果方差的 97%，这证实了是治理策略而非环境初始条件决定了交互的最终走向。

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. 基于角色的分析揭示了两种截然不同的情况：对于那些本身没有转化倾向的访客，治理层是系统能否正常运作的关键；而对于那些已经接近达成一致的访客，原始 LLM 的共情默认设置在很大程度上已经足够。

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. 所有研究结果均基于大模型对大模型的模拟。PID 控制器尚未针对真实人类的不可预测性进行校准，因此在实时流量中验证 EO 是下一步的关键工作。