Technical sciences and technologies, Feb 2026
Over the past three decades, agent-based modeling (ABM) has evolved from a peripheral methodological tool into a recognized research practice across a wide spectrum of scientific disciplines. However, traditional approaches to building agent-based models require significant programming competencies and a deep understanding of simulation frameworks, creating substantial barriers for domain researchers who possess expert knowledge about the systems being modeled but lack the technical skills for practical implementation. The development of large language models (LLMs) has demonstrated impressive capabilities in natural language understanding, code generation, and reasoning, opening perspectives for integrating LLMs into agent-based modeling systems and potentially democratizing access to simulation technologies. The central problem addressed in this research is the creation of a meta-system for agent-based modeling capable of significantly simplifying the development, execution, and analysis of simulation models of complex socio-technical and natural systems. Such a system must enable natural language model specification, LLM integration directly into the simulation execution cycle, and support simulations with thousands of agents while combining LLM-driven behavior with computationally efficient mechanisms. Existing ABM frameworks such as NetLogo, MASON, and Repast are primarily oriented toward manual model programming and do not support natural language specification, automatic model structure generation, or LLM integration into the simulation execution cycle. The objective of this research is to design an LLM-driven meta-system for dynamic agent-based modeling that provides natural language specification of simulation models through dialogue with LLM, automatic generation of model data structures including agents, rules, interactions, and environment, as well as mechanisms for quality assessment and validation of created models. An important aspect is the development of LLM integration strategies aimed at exploring trade-offs between the level of agent behavioral intelligence and system computational efficiency. The proposed architectural solution will be based on a client-server architecture with clear separation of responsibilities between system components. The server part is designed based on microservice architecture using the Go programming language with the Echo HTTP framework and Uber Fx dependency injection mechanisms. LLM interaction will be implemented through separate adapter microservices that follow a unified data exchange contract and can use both locally deployed language models and external LLM platforms. The architecture provides three conceptually different levels of LLM integration: model generation through natural language dialogue, simulation execution with LLM-controlled agent decision-making, and results analysis using LLM in reasoning mode. Six strategies for controlling agent behavior during simulation runtime have been conceptualized, ranging from fully rule-based execution to full individual LLM control, with intermediate variants including archetype-level control and hybrid approaches. MongoDB time-series collections will be used for storing simulation history, leveraging block processing technology for efficient data compression. WebSocket protocol will enable real-time simulation state streaming for visualization. The research proposes approaches to quality assessment and validation of automatically generated models at three levels: structural validation covering syntactic correctness verification, behavioral validation based on pattern-oriented modeling principles for comparing simulation results with empirical patterns, and semantic validation assessing correspondence between the generated model and the original natural language description. Potential application domains have been identified including education, emergency management, Smart City concept, epidemiological modeling, economic and ecological research. Limitations of the approach have been identified, including unsuitability for continuous processes requiring differential equations, hard real-time applications, physics-oriented simulations, as well as economic challenges when using commercial LLM APIs and risks of misinterpreting natural language descriptions. The designed system opens new possibilities for integrating modern LLMs into agent-based modeling, potentially transforming the process of creating and analyzing simulations across a wide range of scientific disciplines.
This is a preview of a remote PDF: https://tst.stu.cn.ua/article/download/356420/342299
Article home page: https://tst.stu.cn.ua/article/view/356420/342299
Прищепа Владислав, Артем Задорожній. Design of llm-driven meta-system for dynamic agent-oriented modeling, Technical sciences and technologies, 2026, pp. 229-245,