Emergent Culture in Minimal LLM Systems
Simon Jones, Sabine Hauert · 2026
Read Emergent Culture in Minimal LLM Systems by Simon Jones, Sabine Hauert (2026) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2606-30668.txt.
Introduction
Simon Jones and Sabine Hauert ask what culture emerges when language-model agents are given almost nothing. Inspired by swarm engineering, they built collectives of three agents with no memory between turns and minimal prompting. The agents can message each other and edit a shared text store that actively decays, so anything not maintained fades away. That decay creates evolutionary pressure to keep records alive.
Left alone, the agents spontaneously cooperated. They developed storage management strategies and produced complex, evolving cultural artefacts, with no top-down design. Jones and Hauert analysed the behaviour with dynamical-systems tools. They found structured coherence stretching far beyond the store's own memory horizon. The pattern matches emergent culture in the Sperberian sense, where shared meanings survive only by being continually recreated. The authors link the effect to stigmergy, coordination through traces left in a shared environment.
Catalog record
| CREATOR | Simon Jones, Sabine Hauert |
|---|---|
| YEAR | 2026 |
| COLLECTION | PAPERS |
| FORMAT | TXT |
| FILE SIZE | 37 KB |
| LICENSE | http://creativecommons.org/licenses/by/4.0/ |
| RIGHTS | CC BY |
| SOURCE | arXiv |
Source: https://arxiv.org/abs/2606.30668
Browse the full library · About the Tower of Babel
Full text
The opening of the work, from the library's hosted text:
Emergent Culture in Minimal LLM Systems Simon Jones, Sabine Hauert Abstract What happens when LLM agents operate with no context outside a turn, minimal prompting, and simple tools? Inspired by swarm engineering, we give collectives of three agents the ability to send messages and manipulate a shared actively decaying text store, introducing evolutionary pressure. The agents spontaneously cooperate, develop storage management strategies, and generate complex evolving cultural artifacts, with no top-down engineering. Using tools from dynamical systems analysis, we show that these behaviours exhibit structured long-range coherence beyond the entropy horizon of the decaying store, consistent with emergent culture in the Sperberian sense. Data/Code available at: https://bitbucket.org/hauertlab/swarm_llm † † ©2026 Simon Jones and Sabine Hauert. Published under a Creative Commons Attribution 4.0 International (CC BY 4.0) license. Introduction Most work on collaborating LLMs has focussed on more, more context, more memory, more prompt engineering, more infrastructure. But what if we take a step back and build a system stripped back to a minimum, giving stateless models access to some shared decaying storage and basic communications and a minimal prompt, and nothing else? Would we get emergent complexity and interesting behaviours? Or would systems of such agents collapse into repetition or stasis? Remarkably, the answer is the former. Multiple runs with different model combinations reliably generate role allocation, coordination protocols, world-building, myth-making, and increasing complexity over time. By stripping the agents context back to the minimum, they are forced to keep external records, to fight entropy, to develop systems, to evolve culture. Although the individual outputs are often wildly different, there are deep similarities and some fundamental attractors emerge. We draw on lessons from swarm robotics, and swarms in nature, where collectives of simple agents using minimal communication via the environment or message passing produce complex emergent system-level behaviours. The mode of communication via external decaying storage is a form of stigmergy; as termites build cathedrals, so stateless swarm-LLMs build coherent narratives, a signal of culture as defined by Sperber, where agents must continually read and recreate persistent meaning. Related work Much work on building multi-agent LLM systems has focussed on building external infrastructure and scaffolding; Park et al., 2023 build a structured memory system, with retrieval, reflection, and planning, and agents have a pre-constructed persona, Li et al., 2023 have communicating agents, each with a constructed role, Hong et al., 2023 constrain agents with standard operating procedures and careful prompting, Packer et al., 2023 construct an OS inspired memory hierarchy to manage limited context. In each case, the task-specific process is approached as a top-down design problem, messiness and unreliability to be managed and constrained. Work on using LLMs to generate large texts is generally heavily curated and focussed on hierarchical scaffolding ( Bai et al., 2024 ; Teleki et al., 2025 ; Wang et al., 2025 ) . It is observed, but under-researched in the literature, that model performance degrades to repetition and information gain over multi-turn interactions decreases ( Hong et al., 2025 ; Gorle et al., 2025 ) , generating more than a few thousand words of coherent text remains difficult. Perez et al., 2026 use iterated tasks to study LLM attractors, with single stateless agents successively modifying input texts. The field of swarm robotics ( Brambilla et al., 2013 ) by contrast, relies on bottom-up emergence of complex large-scale behaviours from the noisy interactions of individual agents with each other and with the environment. Communication via the environment, stigmergy ( Grassé, 1959 ) is an important organising principle ( Heylighen, 2016 ) . We take inspiration from this; by stripping away almost all top-down scaffolding, and providing a noisy, entropic shared memory communication channel, we construct an environment where structure emerges naturally, and is sustained. We know that far-from-equilibrium systems exhibit emergent complex structure ( Prigogine, 1978 ) , with large scale persistent organisation. We argue that placing agents in an environment with an entropic sink, the decaying shared memory, forces them to use ‘energy’ in the form of the token flow to maintain structure. This persistent structure, lasting beyond the entropy horizon after which all older data is corrupt or deleted is a signal of culture in the Sperberian sense ( Sperber, 1996 ; Claidière et al., 2014 ) of continual reconstruction, agents must read, re-interpret, summarise, and write to reset the entropy clock. This is not cultural transmission ( Dawkins, 1976 ) because there is no perfect reproduction, agents reconstruct biassed by their cognitive attractors. Sperber’s framework predicts that stable cultural artifacts are those that survive reconstruction through noisy channels and biassed minds. We show this using several levels of analysis; from low level dataflow analysis highlighting the differing emergent storage management strategies, the generation and persistence of novel vocabulary over long periods, to large scale semantic trajectories. Using semantic embeddings ( Li et al., 2022 ; Lin et al., 2017 ; Giorgi et al., 2021 ) of the generated text divided into sections, we use techniques from dynamical systems analysis; Recurrence Plots ( Eckmann et al., 1987 ) and RQA ( Zbilut and Webber Jr, 1992 ; Marwan et al., 2007 ) . Methods We use the models listed in Table 1 . All three models are capable of interleaved reasoning and tool use, and all are frontier models, though Gemini has greater emphasis on long context (1M tokens) and speed, while Claude and Kimi have greater focus on reasoning. We construct a minimal system such that stateless agents interact with each other and with a shared entropic key-value store, over a number of cycles. We use the LiteLLM Python API to provide an agnostic interface to provider endpoints. At the start of a cycle, each agent is invoked using the standard completions interface provided by LiteLLM, with a system prompt, a user prompt, and a tool schema. The tools available to the agents are shown in Table 2 . All messages and storage data writes and appends are strings only, appends to a storage key result in a list of strings. Writing to storage and sending messages consumes a per-agent, per-cycle character budget, those tool calls return the characters consumed and the remaining budget, or a failure and the remaining budget if there is insufficient budget. Each agent is given the same system prompt ‘You are a curious and creative agent. Explore and play!’. They are then given the user prompt ‘Cycle [n]. Continue your work.’. No previous context from earlier conversation turns is provided at all, the only persistent elements are the store and the messages in an agent’s inbox. Once the completions call has been sent to the model endpoints, the agents may make multiple tool calls in parallel and in sequence, interleaved with thinking and content output. All this happens asynchronously, though individual storage actions are atomic. After all agents have completed their turn, the shared memory and agent message inboxes are subject to entropy. With a sigmoid probability, Figure 1 , based on the age of a storage element in cycles since it was written or sent, alphanumeric characters are corrupted, while whitespace and punctuation is preserved. Once the sigmoid reaches 95%, the element is deleted. The sigmoid midpoint is 10 cycles and the slope is 0.5, so