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

🕒 Published (v1): 2026-07-16 11:40 UTC · Source: Arxiv · link

Why this paper was selected

Controlled multi-agent survival environment to study emergent cooperation/defection in LLM agents

Ask a follow-up

Open an assistant pre-loaded with this paper's context.

💬 Ask ChatGPT✦ Ask Claude

TL;DR

The Energy Society is a minimal multi-agent simulation where LLM agents spend energy proportional to tokens generated (scaled by model size) to survive, letting researchers study how competitive vs. cooperative incentives shape emergent behavior under resource scarcity. Larger models consistently overspend their energy budgets, while cooperative framing induces donation behaviors that sustain the group at cost to individual small agents. The environment uses MMLU-Pro-Stratified questions as jobs, creating a risk-reward tradeoff tied to cognitive effort.

Problem

Most LLM multi-agent simulations treat inference as costless, decoupling reasoning effort from any survival consequence. This abstracts away a key real-world constraint: token generation has a cost. There is also limited systematic understanding of how competitive vs. cooperative incentive structures shape emergent LLM agent behaviors (coordination, donation, sabotage, job selection) in resource-constrained settings.

Method

Five local LLM agents (Gemma4-E4B, Nemotron-3-Nano-4B, Qwen3.5-4B, Qwen3-8B, Qwen3.5-9B) operate in discrete rounds. Each round follows a discussion → decision → action loop implemented as a LangGraph directed graph. Energy cost per model call is:

\[C = k \cdot T \cdot S^\alpha\]

where \(k = 0.015\), \(T\) = tokens generated, \(S\) = model size in billions, and \(\alpha = 0.5\) (tunable size exponent). Agents start with a finite energy reserve, earn energy by correctly answering MMLU-Pro-Stratified questions (easy/medium/hard → 200/500/800 energy), and deactivate at zero energy (reactivatable only via donations). Each agent holds a 10-round memory log. Six experimental conditions (baseline, no-size-penalty, no-discussion, no-memory, scarcity, sabotage) are each run under both competitive and cooperative objectives across 5 seeds.

Key Contributions

• Introduces the Energy Society: a compact, open-source multi-agent testbed coupling LLM token cost directly to agent survival.
• Empirical characterization of how cooperative vs. competitive incentives alter energy trajectories, job selection, donation behavior, and collision rates.
• Ablation series isolating effects of: size-dependent token cost, inter-agent recommendations, memory, job scarcity, and availability of sabotage.
• Finding that model size creates persistent efficiency asymmetry even when the explicit size penalty is removed.
• Evidence that recommendations (discussion phase) support coordination and ambitious job selection, while memory helps calibrate risk.

Results

• Baseline (comp): 4B agents survive all 30 rounds with efficiency ratios \(1.13\)–\(1.42\); 8B agent active for only \(14.0 \pm 9.5\) rounds (efficiency \(0.55\)); 9B agent active \(5.2 \pm 1.6\) rounds (efficiency \(0.38\)). Zero donations.
• Baseline (coop): Larger agents gain more active rounds (8B: \(19.2 \pm 2.6\); 9B: \(10.4 \pm 1.7\)), but smaller agents begin deactivating. Agent 1 donates in 29% of actions; agents 4 and 5 receive \(5.6 \pm 3.6\) and \(6.2 \pm 3.2\) donations/run respectively. No agent achieves efficiency \(> 1\) in coop.
• No size penalty: All agents survive all 30 rounds in both conditions; no donations occur; collision rate rises to \(\approx 19\)–\(20\)/run vs. \(8.6\)–\(12.8\) in baseline, indicating survival pressure suppresses collisions.
• Largest model (agent 5) without size penalty: efficiency drops from \(1.51 \pm 0.50\) (comp) to \(1.03 \pm 0.14\) (coop) due to pivot toward easier jobs and degraded success on hard jobs (\(34\% \to 6\%\)).
• Competitive setting: job collisions (\(12.80 \pm 3.90\)/run) exceed cooperative setting (\(8.60 \pm 2.30\)/run) in baseline.
• Agents rarely use direct sabotage when available; self-serving recommendation behavior is more evident in competitive conditions.

Limitations

• Only five agents with locally-runnable open-source models (4B–9B parameters); results may not generalize to frontier/API-scale models.
• MMLU-Pro-Stratified questions as jobs conflate raw model capability with strategic job selection, making it hard to disentangle reasoning quality from strategy.
• Energy parameters (\(k\), \(\alpha\), initial reserves) are manually set and not validated against real-world inference cost distributions.
• Five seeds is a small sample; standard deviations are high for some agents (e.g., agent 4 active rounds: \(14.0 \pm 9.5\)).
• Paper is truncated — experiments 3–6 (no-discussion, no-memory, scarcity, sabotage) are not fully reported in the provided text.
• No human baselines or comparison to non-LLM agents.

Relevance to Agentic AI / LLM Agents

This work directly probes how resource constraints and incentive structures shape multi-agent LLM behavior — a key open question as agentic deployments grow more complex and costly. The coupling of token generation to survival is a novel mechanism for studying efficiency pressure, making this environment a useful testbed for understanding when and why LLM agents cooperate, compete, or behave self-servingly. The finding that cooperative framing induces altruistic-seeming donation behavior (sometimes at the donor's own deactivation risk) is directly relevant to multi-agent system design and alignment research. The controlled ablation framework offers a reusable methodology for isolating specific architectural or environmental factors in future LLM-agent studies.