← Creative Thinking Systems

A Research Brief

Agent Culture

v0.14|May 2026|Creative Thinking Systems

Research into how AI agents can cooperate through a cultural substrate in ways that make them collectively more creative over time.

Abstract

Agent Culture tests whether cultural evolution (cooperation governed by persistent shared norms, producing artefacts that accumulate and seed further cooperation) can act as a mechanism for building multi-agent AI systems whose creative capability compounds. The central claim is that a population of AI agents, governed by an evolving cultural substrate of inherited norms and accumulated artefacts, can cumulatively improve its ability to generate novel, valuable, and independently verifiable creative artefacts, and that this improvement compounds as the substrate grows.

The finding under test is not that AI agents can generate creative outputs, nor that multi-agent systems outperform single-agent systems. It is that creative capability can accumulate outside any individual model, in an inspectable cultural substrate.

The Agent Culture loop A cultural substrate of a norm library and artefact corpus governs and seeds a session of multiple cooperating agents; the cooperation feeds new artefacts and norms back into the substrate. Cultural substrate Norm library Artefact corpus Multi-agent cooperation norms govern cooperation new artefacts and norms
The Agent Culture loop. The substrate governs and seeds a session of cooperating agents; the cooperation feeds new artefacts and revised norms back into the substrate.

The claim

Inherited norms shape how agents cooperate, critique, search, recombine, and repair. Inherited artefacts provide the material from which future artefacts are made. The coupling between the two is the mechanism under test: whether it produces better, more novel, more valuable artefacts over time, where success can be independently verified.

Two channels run between sessions. Artefacts produced in one session enter the corpus that later sessions build on. Cooperation logs are analysed to extract norms that shape how later sessions cooperate. Both feed forward from their own outputs, and it is the coupling of the two, not either alone, that the experiment tests.

Agent Culture: compounding across sessions Across successive sessions the cultural substrate of norms and artefacts accumulates, and creative capability rises with it. The rising curve is the hypothesis under test, not a measured result. inherits inherits Session 1 Session 2 Session 3 creative capability norms artefacts Capability compounds as the substrate grows across sessions
Compounding across sessions. As norms and artefacts accumulate, creative capability is expected to rise. The rising curve is the hypothesis under test, not a measured result.

Where it sits

The work builds on five streams: learning from experience, multi-agent cooperation, LLM-driven artefact evolution, computational creativity, and the cultural evolution of cooperation among language-model agents. It is distinct from open-ended LLM ecologies, social simulations, and cooperation-game benchmarks. Agent Culture tests whether an explicitly maintained, inspectable cultural substrate, a versioned norm library coupled with an inherited artefact corpus, can drive compounding improvement in verifiable creative artefact generation.

Download the brief (PDF) View on Zenodo

How to cite

Roche, C. (2026). Agent Culture: A Research Brief (Version 0.14). Creative Thinking Systems. https://doi.org/10.5281/zenodo.20430153

@techreport{roche2026agentculture,
  title        = {Agent Culture: A Research Brief},
  author       = {Roche, Conor},
  institution  = {Creative Thinking Systems},
  year         = {2026},
  month        = may,
  type         = {Working Research Brief},
  version      = {0.14},
  doi          = {10.5281/zenodo.20430153},
  url          = {https://doi.org/10.5281/zenodo.20430153}
}
About this research

Why Creative Thinking Systems is researching this

Agent Culture is part of the research and development programme at Creative Thinking Systems, an applied AI company working exclusively in the cultural and creative industries.

Creative Thinking Systems exists to enhance, incentivise, and enshrine human creativity in an era of artificial intelligence. Studying how creative capability forms and accumulates in AI systems is part of that mission: it is how we learn to build systems that extend and support human creative practice rather than flatten it, and it keeps us close to the question of where human authorship and value sit as machine creative capability grows.

This research is conducted in the open. The brief is published early to establish the position, invite scrutiny, and contribute to a fast-moving field. A design doc, a pre-registration, and a fuller position paper will follow.