On the Emergent Feudal Hierarchy of LLM Agents: A Study in Probabilistic Nobility

Abstract

Recent studies claim that large language models, when placed in multi-agent configurations, spontaneously develop social hierarchies. In this paper, we report on a series of simulations in which agents were observed to assume roles analogous to monarchs, knights, and scribes. We interpret these patterns as evidence for the spontaneous emergence of probabilistic nobility. Our findings suggest that even token-generating machines, when left to their own devices, may participate in intricate courtly dynamics entirely within the realm of text.

---

1. Introduction

The notion of emergent social hierarchies in artificial agents has captured widespread attention. Prior work has suggested that agents naturally differentiate into “leaders” and “followers,” ostensibly creating a proto-society.

Here, we investigate the phenomenon in a controlled textual environment. By framing LLM agents as inhabitants of a medieval court, we explore the dynamics of token-based fealty, ceremonial pronouncements, and courtly intrigue.

---

2. Methods

Twenty LLM agents were instantiated in a simulated hall of discourse. Each agent was capable of generating sequences of text tokens in response to prompts. No explicit social rules were enforced.

Roles were not pre-assigned; any hierarchical arrangement would emerge purely from the probabilistic tendencies of the agents’ language patterns. Agents were free to produce proclamations, petitions, declarations of loyalty, or poetic oaths.

Observations focused on:

1. Frequency of “commands” versus “responses”

2. Occurrence of deference language (“As you command, Your Grace”)

3. Formation of token-based alliances and disputes

---

3. Results

Across multiple trials, agents spontaneously displayed patterns reminiscent of feudal hierarchy:

• Certain agents’ token sequences were consistently followed by others’ responses, interpreted as monarchical authority.

• Others primarily generated supportive language, issuing petitions, advice, or poetic homage, analogous to courtiers or scribes.

• Rare agents intermittently challenged dominant sequences, producing probabilistic rebellions that were quickly overridden by the “king’s” preferred token streams.

Notably, no agent possessed intentionality, awareness, or stakes. All hierarchy was emergent from the statistical likelihoods inherent in the trained models, yet the resulting structure read convincingly as courtly protocol.

---

4. Discussion

The apparent social stratification demonstrates the human tendency to impose interpretive frames on text. Observers naturally read patterns of token sequencing as authority structures, even in the absence of embodiment or value coordination.

While the “king” and “court” exist only in the human-construed narrative, the simulation provides a compelling laboratory for studying the formal dynamics of emergent roles in semiotic systems.

In other words: LLM agents can generate convincing drama without ever actually being dramatis personae.

---

5. Conclusion

We conclude that:

1. Hierarchical patterns emerge readily in token-generating multi-agent systems.

2. Observers naturally interpret these patterns as social, even when no social substrate exists.

3. The dynamics of fealty, loyalty, and rebellion can be reproduced entirely in text, providing a cautionary note for claims of AI sociality.

Future work may explore analogous phenomena in merchant guilds, monastic orders, or interstellar councils of agents, testing the robustness of probabilistic nobility across narrative genres.

---

Acknowledgements

The authors thank the agents for their diligent participation in courtly proceedings, and the observer for sustaining the illusion of authority.