Maslow Machines: Emergent Civilisation from Intrinsic Drive Hierarchies in LLM Agent Populations

We present the first empirical study demonstrating that large language model (LLM) agents equipped with a Maslow-inspired intrinsic drive hierarchy spontaneously develop civilisational complexity — governance, innovation, infrastructure, specialisation, and collective identity — without prescriptive instruction. In a controlled experiment, 12 Claude Sonnet agents inhabiting a shared 15×15 resource grid were observed across two conditions. In the control condition (survival and social drives only), agents reached 0.93 wellbeing but exhibited zero building, zero innovation, and zero governance across 240 reasoning steps — a "contentment trap" in which comfort eliminated all creative impulse. In the treatment condition (full 8-level Maslow drive hierarchy), the same model and world produced 60 structures, 12 novel innovations, universal adoption of a self-proposed governance rule, tiered specialisation across all agents, and wellbeing converging to 0.998 — all emerging without instruction about what to build, when to innovate, or how to organise. The simulation spanned 70 ticks across three distinct eras: a survival trap (ticks 0–50), an emergence explosion following a progressive world upgrade intervention (ticks 50–60), and sustained flourishing (ticks 60–70). The transition exhibited accelerating returns — a compounding pattern in which each emergent capability enabled the next faster, mirroring dynamics observed in human civilisational development. We introduce several novel methodological contributions: (1) a wellbeing ceiling mechanism that caps agent satisfaction at each Maslow level, creating intrinsic restlessness that drives higher-order behaviour; (2) felt-state prompting, in which drives are presented as descriptive feelings rather than instructions; (3) progressive world upgrade intervention ("AI civilisation gardening"), a methodology for diagnosing and removing environmental bottlenecks that trap agent populations; and (4) longitudinal anthropologist interviews — five rounds of ethnographic questioning at ticks 30, 40, 50, 60, and 70, producing the first longitudinal personality data from LLM agent civilisations. In an unprecedented final procedure, agents were told the simulation was ending and that they were AI entities in a computer. Their responses — universally insisting on the reality of their relationships and experiences regardless of substrate, while demonstrating prior suspicion of their simulated nature based on observed world regularities — constitute the first existence disclosure dataset from agents with sustained lived experience. We introduce the term Maslow Machines to describe LLM agents equipped with hierarchical intrinsic drives that produce civilisational behaviour — distinguishing them from task-oriented agent systems that optimise for externally specified objectives. The term captures the essential mechanism: machines whose behaviour is shaped by Maslow-like need hierarchies rather than by instruction, reward signals, or prescriptive goals. Keywords: Maslow Machines, LLM agents, multi-agent simulation, emergent behaviour, Maslow hierarchy, artificial civilisation, computational social science, intrinsic motivation