We study the behavior of a heterogeneous pool of agents playing Iterated Prisoner's Dilemma (IPD) on a randomly-generated, dynamically-evolving network. The players include both conventional, rule-based agents (which implement various standard strategies like Tit-for-Tat and Grim Trigger) and also agents backed by Large Language Models (LLMs) which are prompted for decisions. All agents repeatedly play IPD with their neighbors, can change partners over time, and can exchange reputational information about friends-of-friends, possibly truthfully or deceptively. Our goal is to characterize how LLMs behave when embedded in such an environment, measuring how competent they are at the IPD relative to classic rule-based strategies, how they make use of partner choice and reputation, and what their own natural-language rationales reveal about how they conceive of the game environment.

Generative artificial intelligence is increasingly embedded in collaborative settings, and its societal effects depend not only on what it enables individuals to do, but also on how it reshapes interaction itself. This study models generative AI as an interaction intermediary and examines how AI-mediated interaction changes collaboration, trust, and social learning through an agent-based simulation framework. Agents repeatedly choose among peer-first, AI-first, and hybrid interaction modes, and these choices dynamically affect knowledge accumulation, trust updating, tie reinforcement, and collective performance. The results show that AI does not produce a single social outcome; instead, it generates distinct emergent regimes depending on how it is embedded into interaction processes. A complementary and verification-aware use of AI yields the strongest balance between knowledge growth, task quality, inequality reduction, and retention of human interaction. In contrast, AI-first substitution and overreliance weaken the interaction structure that sustains long-run collective learning. The study contributes a compact, mechanism-based framework for understanding when generative AI augments social capability and when it erodes the human interaction patterns on which that capability depends.

The integration of Large Language Models (LLMs) with Agent-Based Modelling (ABM) has attracted growing research attention. While existing work has mainly focused on LLM-empowered agents, this paper contributes to the use of LLMs to support the ABM cycle. Specifically, we apply an LLM-assisted workflow to facilitate model formalisation and implementation in ABM, where stakeholders’ problem understanding is translated into a structured model specification and ultimately executable simulation code. The workflow guides modellers through four iterative steps: extracting model components from problem descriptions, exploring diverse stakeholder perspectives, composing these components into a unified model, and implementing the model in Python. At each step, modellers also collaborate with stakeholders to validate and verify the LLM outputs. Using Claude Sonnet 4.6, we demonstrate the workflow on an electricity market problem. The results show that the workflow maintains consistency throughout the process and produces satisfactory outputs that capture the core problem logic with a modest number of refinement iterations. This workflow not only helps modellers conduct ABM more efficiently, but also lowers barriers to stakeholder participation. Together with our previous work, we demonstrate the workflow’s generality across different LLMs and modelling paradigms. Nevertheless, the inherent output variability of LLMs is discussed, and future work could evaluate the workflow on more complex practical problems.

The growing success of AI systems in producing human-like actions and behaviours has encouraged a search for potential domains of application; agent-based modelling has rapidly emerged as a promising one. The use of AI agents in social simulation has often been normatively justified from the perspective of their increased realism. This text argues that, while this assertion seems hardly contentious at first glance, it requires further elaboration, for it does not fully conform to typical commitments around scientific realism. The analysis claims that a focus on representation and its connection to modelling choices provides a better framework for questions of realism in modelling. Against this backdrop, it is suggested that the idea that AI agents are more realistic has only partial epistemic merit, yet encourages important topics of further research.

Recently, researchers in computer science, riding on the wave of successes of large language models, have turned their attention to employing these models as agents in social simulations. While most of this work is seemingly unaware of the efforts in the social simulation community, and the social simulation community is starting to take notice, bridging the two is seldom the focus of research. In the current paper, we discuss some of the efforts of community bridging and point to promises made and perils noted. Then we turn to more fundamental questions

such as whether large language model agents are appropriate models of human agents. Based on these two analyses we conclude more work is needed, including transdisciplinary work or perhaps we should say cross-community work.