EconoBiome: An LLM-Augmented Agent-Based Model of Post-AI Economic Transitions

EconoBiome is an agent-based model (ABM) of post-AI economic transitions in which large language models (LLMs) serve as the independent experimental variable rather than as simulation tools. The framework comprises eight heterogeneous agent types calibrated to OECD 2025 macroeconomic data, real institutional implementation delays, and a 5x AI acceleration factor estimated from empirical comparisons with historical technology adoption rates. Across 22 experiments and 421 valid simulation runs, we report four main findings. The first two concern the simulation model: within the explored design space, AI market structure — diffuse (open-source) versus oligopolistic — is the strongest predictor of the emergent social regime, producing an 18.9 percentage-point difference in social fracture rates that holds across nine independent parameter perturbations. Second, maximum social risk occurs at intermediate automation rates (ar = 0.45–0.55), not at the highest — a pattern consistent with a vulnerability window hypothesis explained by the gap between technological adoption velocity (factor 6–8x) and institutional response velocity (factor 1.5–2x). The remaining two findings concern LLM behaviour within this simulation environment: Claude Haiku and Claude Sonnet produce statistically identical outcomes (Mann-Whitney U = 50, p = 1.0, n = 10 per model), consistent with shared priors across both tested models, though the underlying mechanism cannot be determined from this experiment alone. Fourth, LLM policy decisions are sensitive to prompt framing: robot tax activation ranges from 0% (market frame) to 100% (neutral frame) under identical economic conditions, while fiscal constraints systematically override framing effects for high-cost policies (chi-square = 45.0, df = 2, p 10% and Gini > 0.30. Code, data (421 simulation runs), and experimental logs are available at: https://github.com/rcarbajalrc/econobiome Conflict of interest: Anthropic (developer of the Claude models used) is explicitly modelled as PlataformaAgent in the simulation. The author has no financial relationship with Anthropic beyond standard API usage fees (~20 EUR total).