Rules measured in play, not stated in isolation, generate the macro dynamics of LLM agent groups

Generative agent-based models use large language models (LLMs) as agents and often calibrate or validate them with questions asked in isolation. Whether such micro-level answers generate the macro-level dynamics of a population is the micro–macro question of agent-based social science. We study it in a repeated public goods game (20 points, MPCR 0.4) with three LLMs (DeepSeek-V4.1-Flash, Mistral-Large-4, GPT-6-Luna) in four studies (1,200 groups), each following a plan written before the data. A simulation built from each agent's stated response to an isolated question mispredicted group trajectories in all three models. A memory-one rule learned from in-game decisions in groups of four (own last contribution, the others' last average, whether someone gave 0, rounds left) reproduced the trajectories out of sample and, frozen in advance, predicted a 15-round game the agents had never played (mean error −0.06 to +0.82 points of 20). In exploratory analyses, the rule also generates the macro structure of four-agent groups: an absorbing collapse state, sticky cooperation, outcomes at the two extremes and, in two models, a tipping point at a single free rider. The stated responses imply averaging dynamics instead. In a pre-registered scaling test, with the forecast deposited publicly before the main run, both models collapsed more often in larger groups: Mistral 0.29 of groups at 4 agents and 0.79 at 16, GPT 0.23 at 4 and 0.69 at 32 (not monotonically in between). Started from each group's first round, the frozen rule predicted Mistral's increase (+0.47 against +0.50 observed; deposited forecast +0.40), partly through cancelling errors, but not GPT's (−0.12 against +0.46; deposited −0.08). Exploratory analyses point to two contributors: agents responded to how many, or what share, of the others defected rather than to whether anyone did (in groups of four, count and share coincide), and in large GPT groups defectors stayed at zero more often. A corrective rule using this information was stopped by a pre-specified gate (improvement 0.047 against a required 0.05); training the original rule on several group sizes helped more (exploratory). Human groups in the identical game (230 groups) occupy a different regime: collapse is escapable and outcomes are spread out. Identity descriptions, varied as a background manipulation, acted through their wording (value sentences and, for one model, an institutional fact) rather than the cities they named. For generative agent-based models the lessons are to measure micro rules in context and at the scales to be simulated. What an LLM agent says in isolation is a poor generator of what its group does, and a rule measured in small groups need not carry over to large ones. Files: preprint_v2.pdf (the preprint), preprint_v2_supplement.pdf (supplement S1–S4), publicgoods_repo_1cf7155.zip (code, written plans with dated amendments, configurations, prompt frames, identity cards, frozen rule tables and forecasts, analysis outputs, figure scripts, test suite, and a Turkish reading copy of the preprint), publicgoods_runs_d71798f.zip (every decision with its reasoning text, run summaries, and raw API logs with the account identifier removed; unchanged since version 1). RELEASE.md in the repository archive gives the contents and the steps to reproduce every number without API calls. The full git history, with the commit dates of the plans, is at https://github.com/gkhngyk/llm-agents-micro-macro. This research was carried out with an AI assistant (Claude) under the author's direction. Version 2 adds sources for the statement in Section 1 that LLM agents are commonly measured with questions asked in isolation (Aher et al., 2023; Argyle et al., 2023; Horton et al., 2023; Park et al., 2024); nothing else changed.