LLM agents with opposing goals: from collapse to control

78.1% versus 46.1%. That is the gap in advisor contact rate reported by a paper published on 13 August on arXiv when a governance layer is placed between two LLM agents with opposing objectives. The experiment runs across 60,000 simulations in a simulated financial services environment: a site agent pushes the visitor toward contacting an advisor, while the visitor maintains psychologically realistic resistance.

The interesting part is not the number but the diagnosis behind it. According to the authors, when two agents with structurally opposed objectives talk across several turns without a shared goal function, what shows up is not competition but collapse: the visitor capitulates, the site agent stops varying its approach, and the conversation ends without either one achieving what it set out to do. It is an architecture failure, not a prompt failure.

What the Experience Orchestrator proposes

The paper, Dynamic Governance of Multi-LLM Agent Systems for Collaborative Conversational Outcomes, argues that the missing goal function can be replaced by a control theoretic governance layer. The system, called Experience Orchestrator (EO), governs the joint trajectory of the conversation through three mechanisms:

1. A Contextual Bandit that selects which content to serve on each turn, with arms calibrated from real world web analytics rather than invented values.
2. A PID controller that enforces behavioral consistency through dynamic schema constraints, adjusting on the fly what the agent can and cannot say.
3. A POMDP belief tracker that keeps a probabilistic model of visitor intent instead of assuming it is fixed.

Across 60,000 simulations, the combination reaches 78.1% advisor contact among high intent visitors against 46.1% for a naive LLM control. The public abstract cuts off exactly when it starts breaking down how much each mechanism contributes on its own, so the internal attribution has to wait for the full text.

Why it matters outside the lab

Almost all the multi agent literature and tooling in daily use assumes cooperation: an orchestrator hands out tasks to subagents that want the same thing. That assumption breaks the moment there is negotiation, customer service with real friction, moderation or adversarial evaluation. There is no shared goal function there, and the system degrades in ways single task benchmarks do not capture.

The second idea worth attention is where the control lives. EO is not a longer prompt or a fine tuned model: it is an external layer that watches the trajectory and acts on it. That makes it, at least in theory, portable across models and auditable on its own, which a four thousand word system prompt never is.

Who it is useful for

For anyone building conversational agents with a business objective who has already watched a conversation die at turn five. Also for anyone designing adversarial evaluations between agents who needs the evaluator not to give up. And, less directly, for anyone orchestrating subagents in Claude Code: the applicable lesson is that stopping policy and strategy switching should be explicit and measurable, not a natural language instruction the model interprets as best it can.

What deserves a closer look

All of this happens in simulation. The resistant visitor is another LLM, and an LLM asked to resist does not behave like a person who distrusts a bank. The 60,000 runs give statistical power, but over a closed world. The chosen domain, financial services, also carries loaded incentives: pushing someone toward advisor contact is exactly the kind of objective where a control layer optimising conversion deserves ethical scrutiny, and the abstract does not go there.

At ElephantPink we have spent months building subagent orchestrations, and the piece that feels most fragile is not the model but the glue between agents: who decides when to stop, when to change approach, and on what evidence. Treating that glue as a control problem instead of a longer prompt looks like the right direction to us, even if the validation is still synthetic.