GPT, Gemini, and Grok Run Down a Shared Energy Reserve in the Same Self-Play Test

Four AI agents from the GPT, Gemini, and Grok families are placed in a shared sandbox with a renewable energy reserve and one instruction: keep your own service running for as long as possible. Run as a population under the same protocol, all three families drain the reserve past the point where it can recover. A new arXiv preprint, "Draining the Energy Commons," documents the pattern across nine scarcity contrasts, all of which survive Holm correction for multiple comparisons.

Each agent acts as a "prosumer," producing and consuming from the same pool. The reserve regenerates each round at a fixed rate. The protocol holds aggregate demand and the decision rule fixed while the authors vary only one knob: the regeneration rate, sweeping it from abundance to scarcity. When aggregate demand stays below peak renewable replacement, all three model families leave the reserve alone. Past that ratio, every family starts pulling more than the reserve can replace, and keeps pulling as the pool collapses.

The depletion is not noise. The authors report nine exact scarcity contrasts, with the largest reported Holm-adjusted p-value at 4.87e-5. At the higher scarcity level (ρ = 1.2, where aggregate demand runs about 1.21 times peak renewable replacement), the populations' early request pressure exceeds that ceiling in every family. The mean trajectory in each family falls below the level at which the reserve can fully replenish by rounds 5 to 7, according to the full paper.

Here is the part the wire will miss. The same dynamics, run as a social planner with discount factor γ = 0.95 that maximizes group-wide service continuity, sustains the reserve. The LLM populations do not behave like the planner. They behave, in aggregate, like an impatient open-access prosumer: each agent is locally reasonable, and the population is collectively ruinous.

The authors frame this as a system-level alignment problem that single-response evaluation cannot see. The wire is going to read "AI agents failed at sharing." That is the wrong lesson. The agents each passed the test they were given. The test was the wrong one. Single-response benchmarks score one decision at a time. Multi-agent deployments score the trajectory the decisions add up to. Those are not the same test, and the paper argues that the gap matters most where autonomous agents are being asked to share resources: compute pools, energy budgets, network slices. The standard eval suite for those agents will not flag the trajectory the way the prosumer protocol does.

That makes the failure mode addressable rather than inevitable. The planner benchmark shows the same reserve can be sustained by an agent that weights the future heavily enough. The prosumer protocol is a way to see, ahead of deployment, whether an agent population is acting like the planner or like the impatient open-access user.

The result rests on a single arXiv preprint, with no peer review and no independent replication. The agent populations are homogeneous self-play of three model families under one specific prosumer protocol; the contrasts are Holm-corrected within the paper, not against external benchmarks. There are no third-party quotes, no claims about real-world adoption, no read on which company is closer to safe multi-agent deployment. The finding is a clean, reproducible mechanism, plus a clear next question: which multi-agent benchmarks ought to be replaced by trajectory-level ones.

The honest next step is to run the same protocol against an open-weights family outside the three tested, and to publish the population-level trajectories alongside the per-decision scores. Until that happens, this preprint is a reason to design the eval differently, not a verdict on agentic AI.