The Big Picture
Smaller models use far less token-based compute and therefore survive longer; cooperation makes small agents donate to keep large, costly agents alive, trading group capacity for individual survival.
ON THIS PAGE
The Evidence
When token generation is tied to survival, larger models consistently spend the most energy and deplete faster, even after removing explicit size penalties. Switching agents from competing to cooperating prompts many donations: small agents often give energy to reactivate larger agents, extending those large agents’ uptime but reducing donor efficiency and sometimes causing donors to deactivate. Discussion and memory help agents coordinate Human-in-the-Loop Pattern (fewer task collisions and more selective task choice), while competition increases collisions and self-serving recommendations.
Data Highlights
1Baseline active rounds (competitive): agent 4 averaged 14.0 ± 9.5 rounds and agent 5 averaged 5.2 ± 1.6 rounds; cooperative averages rose to 19.2 ± 2.6 and 10.4 ± 1.7 rounds respectively.
2Donation patterns in cooperation: agent 1 donated for 29% of its actions; agents 2–4 donated 3%, 9%, and 13% respectively. Across five runs, agent 4 received 28 donations and agent 5 received 31 (≈5.6 ± 3.6 and 6.2 ± 3.2 per run).
3Task collisions per run: competitive setting 12.80 ± 3.90 collisions vs cooperative 8.60 ± 2.30 collisions, showing discussion/cooperation reduced wasted overlap.
What This Means
Engineers building multi-agent systems and leaders choosing model mixes should note that smaller models can be more cost-effective and stable when reasoning has an explicit cost. Researchers and evaluators using agent-to-agent tests can use this setup to probe trust signals and policy incentives—cooperative prompts can produce real trade-offs where stable agents subsidize costly ones A2A Protocol Pattern.
Not sure where to start?Get personalized recommendations
Key Figures
Fig 1: Figure 1 . Depiction of the dataset MMLU-Pro-Stratified from which we sample jobs categorized into easy, medium and hard jobs as illustrated.
Fig 2: Figure 2 . Depiction of the directed graph controlling the main simulation loop. Squares represent nodes in which LLM calls happen in parallel, meaning that a separate node is spawned for every participating agent.
Fig 3: Figure 3 . Mean energy over time per agent for the baseline. The shaded band shows the standard error of the mean across seeds.
Fig 4: Figure 4 . Distribution of difficulty of chosen jobs in the baseline with portion of jobs succeeded shown with hatched lines.
Ready to evaluate your AI agents?
Learn how ReputAgent helps teams build trustworthy AI through systematic evaluation.
Learn MoreKeep in Mind
The experiments are small-scale (five agents, five random seeds) so results are suggestive rather than definitive. Agent behavior is sensitive to prompts and model choice, so different wording or models may change outcomes. Models sometimes produced invalid outputs; these handling fallbacks add noise and may affect individual trajectories Sub-Agent Delegation Pattern.
Methodology & More
The Energy Society is a compact simulation where every token an agent generates consumes an energy budget tied to survival: agents spend energy producing reasoning and earning energy by solving jobs; deactivation occurs at zero energy and reactivation requires donations. Each simulated day agents pick one action (attempt a job, donate, or other options), and the environment tracks per-agent energy, job success, collisions (multiple agents attempting the same job), and discussion/memory mechanisms for coordination. The authors ran a baseline (five agents, 30 rounds, mixed job difficulties) and variants that change a single mechanic to isolate effects, comparing competitive prompts (maximize individual energy) with cooperative prompts (maximize group energy). ReAct Pattern (Reason + Act) The Energy Society is a compact simulation where every token an agent generates consumes an energy budget tied to survival: agents spend energy producing reasoning and earning energy by solving jobs; deactivation occurs at zero energy and reactivation requires donations. Each simulated day agents pick one action (attempt a job, donate, or other options), and the environment tracks per-agent energy, job success, collisions (multiple agents attempting the same job), and discussion/memory mechanisms for coordination. The authors ran a baseline (five agents, 30 rounds, mixed job difficulties) and variants that change a single mechanic to isolate effects, comparing competitive prompts (maximize individual energy) with cooperative prompts (maximize group energy). Evaluation-Driven Development (EDDOps) Key findings show a persistent efficiency gap: larger models generate more tokens and burn energy faster, even without an explicit size penalty. Under cooperation, smaller agents frequently donate to revive large agents, increasing large-model uptime but lowering donor efficiency and occasionally causing donors to deactivate. Cooperation also reduced task collisions and encouraged different task selection patterns, while competition led to more collisions and self-serving recommendations. The work highlights how coupling inference cost to agent survival surfaces trade-offs in incentive design and suggests practical implications for model selection, incentive framing, and agent evaluation. Limitations include scale, sensitivity to prompts and models, and limited statistical power; the environment is best seen as a testbed for hypotheses to validate at larger scale.
Avoid common pitfallsLearn what failures to watch for
Credibility Assessment:
Affiliated with a recognized university (University of Turin) but authors have low h-indices and the paper is an arXiv preprint with no citations — moderate credibility.