Governance layer keeps competing LLM agents from collapsing conversations

A new arXiv paper argues that conversations between LLM agents with opposing goals can fail without a shared control layer.

The researchers study a simulated financial services setting where one agent tries to guide a visitor toward advisor contact while the visitor maintains realistic resistance. Without coordination, the paper says the visitor tends to capitulate, the site agent stops adapting, and the exchange ends without either side’s stated objective being handled well.

Their proposed Experience Orchestrator combines three mechanisms: a contextual bandit to choose content options, a PID controller to enforce behavioral consistency through schema constraints, and a partially observable Markov decision process belief tracker to estimate visitor intent.

Across 60,000 simulations, the system raised high-intent advisor contact rates from 46.1% to 78.1% compared with a naive LLM control. The result is still simulation-based, but it points to a larger design lesson for agent systems: orchestration and measurement may matter as much as the individual model prompts.