The Energy Society: A Simulation Environment for Studying Agent Cooperation under Survival Pressure

This paper investigates how Large Language Model (LLM) agents behave when their survival is directly tied to the computational cost of their own inference. By introducing a simulation environment called The Energy Society, the authors explore how competitive versus cooperative incentives shape emergent social dynamics, such as job selection, resource redistribution, and coordination, under strict survival pressure.

The researchers created a multi-agent simulation where agents must manage a finite energy reserve. Energy is consumed based on the number of tokens generated (scaled by model size) and replenished by successfully completing tasks (jobs) or receiving donations from other agents. If an agent's energy reaches zero, it deactivates. The study compares two primary conditions: a competitive setting, where agents aim to maximize their own energy, and a cooperative setting, where they aim to maximize the group's total energy. The authors systematically varied environmental factors, such as the presence of memory, the ability to recommend actions, and the existence of a size-dependent token cost, to isolate the drivers of agent behavior.

The study reveals that model size creates a persistent efficiency asymmetry; larger models consume energy faster than they can replenish it, even when the size-dependent cost is removed. In competitive settings, larger agents frequently deactivate, while smaller agents demonstrate unexpected resilience. When cooperative incentives are introduced, agents shift their behavior to support the group, frequently donating energy to reactivate deactivated peers, sometimes even at the cost of their own survival. Furthermore, the ability to recommend actions to others facilitates better coordination, and memory allows agents to calibrate their risk-taking when selecting jobs of varying difficulty.