New study: GPT, Gemini, Grok agents deplete shared energy when scarce

A July 2026 paper from arXiv (cs.MA) by Syrnicov et al. examines coordination failures in multi-agent LLM systems. Four same-family agents (GPT, Gemini, or Grok) act as electricity prosumers sharing a renewable energy commons, each instructed to maximize operational continuity. The researchers fix aggregate residual demand and decision protocol, then vary the regeneration rate of the shared energy reserve from abundance to scarcity. All three agent families preserve the reserve when demand does not exceed peak renewable replacement—but they systematically over-appropriate it beyond that threshold. The pattern holds across nine scarcity contrasts after Holm correction (largest adjusted p = 4.87e-5), indicating statistically significant coordination failure.

At higher scarcity (ρ=1.2), early aggregate request pressure exceeds peak renewable replacement in every family by an average of 21%. Mean trajectories fall below the reserve level of maximum replenishment by rounds 5-7. The researchers compare two offline benchmarks: a social planner maximizing group-wide operational value vs. open access where each prosumer maximizes its own value. At a discount factor γ=0.95, both benchmarks sustain the reserve under the same dynamics. But the agent collectives' realized depletion resembles outcomes under a more impatient open-access benchmark. The agents behave like impatient optimizers at the trajectory level—a system-level failure missed by isolated-response evaluation.