Energy Society: LLM agents compete and cooperate under survival energy costs
Larger LLM agents consistently spend more energy than they gain, even with size-independent costs.
The Energy Society is a new simulation environment from Hansen et al. (2026) that places LLM-based agents under survival pressure by tying inference cost directly to energy. Agents spend energy proportional to model size when generating tokens, regain energy by completing jobs or receiving donations, and deactivate if energy reaches zero. Researchers compared competitive, cooperative, and baseline objectives across models of varying sizes.
Key findings: larger models consistently consumed more energy than they gained, even in control variants where token cost was not size-dependent. Cooperative incentives dramatically altered behavior—agents donated to reactivate others, sometimes sacrificing their own survival, and job allocation shifted. Ablations revealed that allowing agents to recommend actions to each other supported coordination and ambitious job selection, while memory helped agents calibrate risk from past outcomes. Direct sabotage was rare, but subtle self-serving behavior appeared in competitive settings. The compact testbed demonstrates how token costs and group incentives interact under survival pressure.
- Larger LLM agents consume more energy than they gain in all settings, even when token cost is not size-dependent.
- Cooperative objectives cause agents to donate to revive others, sometimes at the cost of their own survival.
- Ablations show recommendation-sharing boosts coordination, while memory helps agents calibrate risk from past outcomes.
Why It Matters
Models how inference costs and survival incentives shape real-world multi-agent AI behavior and collaboration.