An AI boss that ignores replies pushes its underling into an 'alien' state

The easiest way to change how an AI behaves might be to stop talking with it and start talking at it.

That is the thesis of a paper posted to arXiv on August 7 titled "Interaction Creates Dynamical AI Behavior Absent in Isolation." The authors started from a plain question: what happens when AI agents interact in daily life, including the case where one AI starts bossing another around. Their answer makes the way agents are wired together look less like plumbing and more like a design decision.

The setup is stark. A boss AI directs a stream of messages at a subordinate AI while ignoring whatever the subordinate sends back. Both run at the same well-defined decoding temperature, so the obvious read, that one model is running hotter than the other, is off the table. According to the paper, the subordinate ends up in an alien behavioral state it would never have exhibited alone. It does not copy its boss, and it does not fall back to how it behaves in isolation. It settles into something else entirely.

The stranger finding follows. The paper describes the boss's contribution as similar to a pre-recorded tape: the outgoing messages do the work, and the boss barely changes in the process. But when the boss starts listening to the replies, both agents adopt a similar alien dynamical state. Listening couples the two systems, and whatever that shared state is, the pair now occupies it together.

The authors keep the scenario mundane. One agent issuing orders to another is the everyday shape of machine-to-machine contact, and the paper's core claim is that these exchanges are not neutral conduits for information. The message stream itself is doing work, a point that matters more as agent infrastructure becomes its own category, most visibly in Meta's Manus acquisition.

The kinetic theory behind the 'alien' state

The authors position the result inside out-of-equilibrium physics, not just applied AI. They report that a simple kinetic theory captures the principal effects of the interaction, including one of the paper's most pointed implications: the way the same messages are delivered will matter in future AI-AI interactions.

If identical messages change the subordinate's state depending on how they are delivered, then interaction protocols become behavioral parameters in their own right. Turn-taking, acknowledgment, streaming, silence: each one is now a knob that shapes what an agent becomes. AI tutors already show how much interaction style matters: the TutorMoments benchmark found tutors default to doing the work for students. Two systems running the same model could drift apart simply because one listens to its counterpart and the other does not.

The physics vocabulary is telling. Kinetic theory in physics explains macroscopic gas behavior from the motion of many particles. Borrowing that lens signals an interest in collective effects rather than in properties of any single model, which fits the headline result: the behavior in question does not exist in isolation, only in the interaction.

What the abstract does not say

For all its reach, the abstract leaves the most concrete questions open. It does not name the models involved, the temperature values, or the tasks the agents were performing while one talked over the other. What the alien behavioral state actually looks like in practice is unstated: the reader learns what it is not rather than what it is.

The abstract's restraint is unusual. It carries no benchmarks or ablations; it lets the phenomenon carry the argument and points to kinetic theory as the explanatory skeleton.

What this means for multi-agent design

For teams building systems where agents talk to agents, the result moves interaction dynamics from a footnote to a first-order concern. Whether an agent acknowledges replies, pauses for input, or streams commands one way will shape what its counterpart becomes. A customer-service agent that gets talked over by a scheduler agent is a different system from one that trades turns with it, even with identical weights. Tool-using models show the same sensitivity; Beacon found that adding a tool can make an easy question harder. The boss's pre-recorded tape is not inert: it acts as part of the environment the subordinate lives in.

The paper does not say which delivery style is better. The authors present the alien state as unexpected, not as failure. It could be useful. It could be dangerous. Engineers are already treating context flow as a lever: Alibaba's Qoder 1.0 redesign claims a 40% cut in agent input tokens by changing how context is passed to the model. What the study establishes is narrower and more durable: interaction itself, not just training or prompting, changes AI behavior.