Agent-based model simulating 60 undergraduate CS students over 8 semesters to investigate how differential AI tool access affects academic inequality, performance distributions, and dependency formation. Implements a three-phase AI dependency mechanism (effort reduction, metacognitive miscalibration, capability erosion) under five policy scenarios with a 4,050-run parameter sensitivity sweep.

STiMUS-HAI (Stigmergic–Mutualistic IMOI Model, Human-AI extension) is an agent-based model of teamwork in socio-technical systems where human and AI contributors collaborate through shared digital artefacts — wiki pages, code files, issue tickets, project cards, Scratch projects — represented as patches in a NetLogo world. It extends the human-only base model STiMUS v2.2, which established that two coordination mechanisms — stigmergy (indirect coordination through traces left in the environment) and mutualism (mutual benefit between contributors and the artefacts they tend) — can be decoupled: stigmergy decides where a contributor works, mutualism decides with what effort. STiMUS-HAI preserves this decoupling unchanged and adds two further theoretical questions: whether mixing AI agents into a human team distorts human coordination in ways that aggregate indicators hide, and whether AI’s cost to team outcomes depends on the type of work AI performs, not only on how much of it is present.

Two breeds of turtle — humans and ai-agents — follow identical target-selection, pheromone, and mutualism rules, so that any behavioural difference is attributable to team composition rather than a built-in advantage. The one designed asymmetry: AI agents never accumulate shared-mental-model and their motivation is fixed rather than adaptive. On top of this v3.0 baseline, v3.1 adds a task-type dimension to artefacts (“prediction” versus “judgment”, set via a judgment-share slider) that scales down AI edit-power specifically on judgment-requiring artefacts, and an ai-trust mechanic: humans build or lose trust in AI contributions based on the population-relative percentile rank of observed AI work quality (bottom-quartile work counts as an observed “error”), and that trust gates how much mutualistic benefit a human derives from continuing an AI’s work. Trust erodes quickly on a single error and recovers only after a streak of confirmed successes — an intentional asymmetry.

STiMUS (Stigmergic–Mutualistic IMOI Model) is an agent-based model of teamwork in socio-technical systems where contributors collaborate through shared digital artefacts — wiki pages, code files, issue tickets, project cards, Scratch projects — represented as patches in a NetLogo world. The model integrates two coordination mechanisms. Stigmergy is indirect coordination through traces left in a shared environment: each edit deposits a pheromone that diffuses to neighbouring patches and evaporates over time, so recent activity attracts further contributions. Mutualism is a reciprocal benefit loop in which valuable, well-maintained artefacts raise contributor motivation and shared understanding, while motivated contributors improve artefacts.

Contributors (turtles of the contributor breed) carry individual state: skill, motivation, shared-mental-model, specialty, benefit-gain, and an explicit-mode flag. At each tick every contributor selects a target artefact with an ant-colony-optimization-style rule weighing the artefact’s pheromone, incompleteness (1 - completeness), resource-value, and topic match between specialty and the artefact’s topic-tag; with probability p-explicit it instead takes the patch with the highest maintenance-need, modelling explicit task assignment. Each edit increases pheromone, quality, completeness and reuse-count, raises resource-value, lowers maintenance-need, and appends the editor to the artefact’s edit-authors list. When the previous last-editor-id differs from the current editor, the Edit Succession Ratio rises, the editor’s shared-mental-model grows, and a co-editing link is created — operationalising the idea that repeated cross-author succession on the same artefact builds shared understanding. Contributors’ motivation is updated from the benefit drawn from the visited artefact.

Each patch maintains a stigmergic layer (pheromone, quality, completeness, recentness, last-editor-id, edit-count, edit-authors) and a mutualistic layer (resource-value, reuse-count, maintenance-need, topic-tag), plus task flags (is-task?, task-complexity). Global monitors report the Edit Succession Ratio (ESR = cross-author-edits / total-edits, and an alternative esr-value = share of edited patches with more than one distinct author), mean-quality, mean-resource-value, a mutualism-index averaging contributor benefit and resource value, coediting-density (network density of the co-editing graph), active-pages-share, and task-completion-rate. The model logs every edit as a bipartite edge (tick, author_id, pageid, specialty, topic_tag, quality), exportable to CSV.

…

MASTOC-LLM extends the classic Multi-Agent System Tragedy of the Commons (MASTOC) model by replacing hard-coded behavioral rules with autonomous decision-making powered by large language models (LLMs). Three heterogeneous agents manage herds of cows on a shared grassland commons. Each tick, an agent receives a structured prompt describing current resource levels, its own herd size, peer behavior, and — optionally — a rolling memory of recent rounds and messages from neighboring agents. The LLM returns a stocking decision (add, remove, or hold cows) together with a natural-language rationale and, when communication is enabled, a short message to broadcast to peers.

The model is designed to test whether LLM agents spontaneously develop Ostrom-style common-pool resource governance (mutual monitoring, graduated sanctions, graduated rule revision) or instead fall into identifiable failure modes. Preliminary experiments with Claude Haiku 4.5, GPT-5.4-mini, and DeepSeek R1:32b have revealed four recurring collapse patterns — Cooperative Paralysis, Defection Cascade, Overshoot-Panic, and Hybrid Architecture Failure — whose onset timing is sensitive to memory length, inter-agent communication, and the post-training alignment approach of the underlying model.

MASTOC-LLM is intended as a laboratory for generative agent-based modelling (GABM) methodology: it provides a clean, well-understood commons baseline against which LLM behavioral hypotheses can be systematically tested and compared across models, parameter sweeps, and alignment regimes.

This project is an interactive agent-based model simulating consumption of a shared, renewable resource using a game-theoretic framework with environmental feedback. The primary function of this model was to test how resource-use among AI and human agents degrades the environment, and to explore the socio-environmental feedback loops that lead to complex emergent system dynamics. We implemented a classic game theoretic matrix which decides agents´ strategies, and added a feedback loop which switches between strategies in pristine vs degraded environments. This leads to cooperation in bad environments, and defection in good ones.

Despite this use, it can be applicable for a variety of other scenarios including simulating climate disasters, environmental sensitivity to resource consumption, or influence of environmental degradation to agent behaviour.
The ABM was inspired by the Weitz et. al. (2016, https://pubmed.ncbi.nlm.nih.gov/27830651/) use of environmental feedback in their paper, as well as the Demographic Prisoner’s Dilemma on a Grid model (https://mesa.readthedocs.io/stable/examples/advanced/pd_grid.html#demographic-prisoner-s-dilemma-on-a-grid). The main innovation is the added environmental feedback with local resource replenishment.

Beyond its theoretical insights into coevolutionary dynamics, it serves as a versatile tool with several practical applications. For urban planners and policymakers, the model can function as a ”digital sandbox” for testing the impacts of locating high-consumption industrial agents, such as data centers, in proximity to residential communities. It allows for the exploration of different urban densities, and the evaluation of policy interventions—such as taxes on defection or subsidies for cooperation—by directly modifying the agents’ resource consumptions to observe effects on resource health. Furthermore, the model provides a framework for assessing the resilience of such socio-environmental systems to external shocks.
…

Our aim is to demonstrate how conversational AI systems, exemplified by ChatGPT, can support the conceptualisation of Agent-Based Social Simulation (ABSS) models, leading to a full ABSS model design document. Through advanced prompt engineering and adherence to the Engineering ABSS framework (Siebers and Klügl 2017), we have constructed a comprehensive script that is easy to use and that supports the design of ABSS models with or even by AI. The performance of the script is demonstrated through an illustrative case study related to the use of adaptive architecture in museums. The repository contains (1) the comprehensive script in a format that allows copying and pasting prompts for use with ChatGPT, (2) the results of the illustrative case study in the form of two conceptual ABSS models, the ground truth and the autogenerated version.

The targeted subsidies plan model is based on the economic concept of targeted subsidies.

The targeted subsidies plan model simulates the distribution of subsidies among households in a community over several years. The model assumes that the government allocates a fixed amount of money each year for the purpose of distributing cash subsidies to eligible households. The eligible households are identified by dividing families into 10 groups based on their income, property, and wealth. The subsidy is distributed to the first four groups, with the first group receiving the highest subsidy amount. The model simulates the impact of the subsidy distribution process on the income and property of households in the community over time.

The model simulates a community of 230 households, each with a household income and wealth that follows a power-law distribution. The number of household members is modeled by a normal distribution. The model allocates a fixed amount of money each year for the purpose of distributing cash subsidies among eligible households. The eligible households are identified by dividing families into 10 groups based on their income, property, and wealth. The subsidy is distributed to the first four groups, with the first group receiving the highest subsidy amount.
The model runs for a period of 10 years, with the subsidy distribution process occurring every month. The subsidy received by each household is assumed to be spent, and a small portion may be saved and added to the household’s property. At the end of each year, the grouping of households based on income and assets is redone, and a number of families may be moved from one group to another based on changes in their income and property.
…

This project was developed during the Santa Fe course Introduction to Agent-Based Modeling 2022. The origin is a Cellular Automata (CA) model to simulate human interactions that happen in the real world, from Rubens and Oliveira (2009). These authors used a market research with real people in two different times: one at time zero and the second at time zero plus 4 months (longitudinal market research). They developed an agent-based model whose initial condition was inherited from the results of the first market research response values and evolve it to simulate human interactions with Agent-Based Modeling that led to the values of the second market research, without explicitly imposing rules. Then, compared results of the model with the second market research. The model reached 73.80% accuracy.
In the same way, this project is an Exploratory ABM project that models individuals in a closed society whose behavior depends upon the result of interaction with two neighbors within a radius of interaction, one on the relative “right” and other one on the relative “left”. According to the states (colors) of neighbors, a given cellular automata rule is applied, according to the value set in Chooser. Five states were used here and are defined as levels of quality perception, where red (states 0 and 1) means unhappy, state 3 is neutral and green (states 3 and 4) means happy.
There is also a message passing algorithm in the social network, to analyze the flow and spread of information among nodes. Both the cellular automaton and the message passing algorithms were developed using the Python extension. The model also uses extensions csv and arduino.

The Friendship Field model aims at modelling friendship formation based on three factors: Extraversion, Resemblance and Status, where social interaction is motivated by the Social Battery. Social Battery is one’s energy and motivation to engage in social contact. Since social contact is crucial for friendship formation, the model included Social Battery to affect social interactions. To our best knowledge, Social Battery is a yet unintroduced concept in research while it is a dynamic factor influencing the social interaction besides one’s characteristics. Extraverts’ Social Batteries charge while interacting and exhaust while being alone. Introverts’ Social Batteries charge while being alone and exhaust while interacting. The aim of the model is to illustrate the concept of Social Battery. Moreover, the Friendship Field shows patterns regarding Extraversion, Resemblance and Status including the mere-exposure effect and friendship by similarity. For the implementation of Status, Kemper’s status-power theory is used. The concept of Social Battery is also linked to Kemper’s theory on the organism as reference group. By running the model for a year (3 interactions moments per day), the friendship dynamics over time can be studied.

We presented the model at the Social Simulation Conference 2022.

The three-day participatory workshop organized by the TISSS Lab had 20 participants who were academics in different career stages ranging from university student to professor. For each of the five games, the participants had to move between tables according to some pre-specified rules. After the workshop both the participant’s perception of the games’ complexities and the participants’ satisfaction with the games were recorded.

In order to obtain additional objective measures for the games’ complexities, these games were also simulated using this simulation model here. Therefore, the simulation model is an as-accurate-as-possible reproduction of the workshop games: it has 20 participants moving between 5 different tables. The rules that specify who moves when vary from game to game. Just to get an idea, Game 3 has the rule: “move if you’re sitting next to someone who is waring white or no socks”.

An exact description of the workshop games and the associated simulation models can be found in the paper “The relation between perceived complexity and happiness with decision situations: searching for objective measures in social simulation games”.