LLM-ABM Methodological Handbook

Agent-based modeling has transformed social science. The arrival of large language models changed it again. LLM-powered agents exhibit richer, more context-sensitive behavior than any rule-based system could encode. Yet three hard problems remain: how to reproduce results when commercial APIs update silently, how to manage the engineering and cost of thousands of LLM calls per experiment, and how to demonstrate that an LLM agent behaves like a real human rather than merely producing plausible text. This is the first systematic, end-to-end methodological guide for the LLM-ABM era. It takes readers from research question to publication, with concrete checklists, worked examples, exercises, and curated reading in every chapter. Five parts: Part I (Foundations) establishes the complexity-science basis of ABM. Part II (Model Construction) covers platform selection, code implementation, parameterization, and initialization. Part III (Verification and Analysis) addresses verification, sensitivity analysis, and results interpretation. Part IV (LLM Agents in ABM), expanded in this edition from three chapters to seven, delivers the book's core: agent design, memory architecture, prompt engineering, cost control, population construction, evaluation benchmarks, and research ethics. Part V (Open Science and Publishing) tackles reproducibility protocols, cost disclosure, and peer review. A distinctive contribution is the ODD-LLM extension, a proposed addition to the standard ODD protocol for documenting LLM-specific configuration: model version, prompt templates, memory architecture, and sampling parameters. Without such documentation, LLM-ABM results cannot be reproduced. Seven appendices provide ready-to-use instruments: a complete ODD/ODD-LLM template, a platform comparison table, a sensitivity analysis quick reference, an annotated reading list, an end-to-end workflow checklist, a glossary, and a troubleshooting guide. For graduate students, researchers, and course instructors in computational social science, this book is both a practical manual and a methodological reference for rigorous, reproducible agent-based research in the age of large language models.