Groups of AI agents underuse resources despite shared evidence of increased capacity

Takahiro Ezaki, Naoto Imura, Katsuhiro NishinariarXiv 2026 · arXiv

TL;DR

Even when AI agents are perfectly informed about resource availability, they often fail to use that capacity efficiently because they all react to the same information in the same way. This "algorithmic monoculture" leads to crowding on recovered resources while leaving others idle, a problem that can be mitigated by giving agents distinct, private identifiers to break their decision-making symmetry.

What Problem Does This Paper Address?

As AI agents increasingly manage tasks like compute job submission or infrastructure reservations, they must share limited resources. The goal is collective efficiency—completing as many jobs as possible. However, agents often rely on delayed status reports (e.g., a "service status" page that is two minutes behind).

The authors identify a three-part challenge for efficient populations:

  1. Discovery: Recognizing that a failed resource has recovered before the official report says so.
  2. Action: Acting on evidence of that recovery.
  3. Distribution: Spreading demand across all available resources to avoid oversubscribing one while others sit empty.

The core problem is that if all agents are "well-informed" with the same shared signal, they may all rush to the same resource simultaneously, creating a "stampede" that wastes total system capacity.

Methods

The researchers designed a simulation with 16 AI agents, each trying to complete one job across two servers (A and B) within a four-round deadline.

  • The Scenario: Server A starts at zero capacity and recovers to a capacity of six in round 2. Server B has a constant capacity of two.
  • The Information Gap: The shared capacity report is delayed by two rounds, meaning agents see Server A as "unavailable" even after it has recovered.
  • The Intervention: The authors tested different prompts, including "recovery descriptions" (telling agents A might recover) and "success records" (showing proof that another agent successfully used A).
  • Symmetry Breaking: To address the distribution problem, they introduced a private_draw—a random number given to each agent. They compared groups where every agent got the same number (Condition C) versus groups where each agent got a unique number (Condition I).

The study evaluated five large language models (LLMs): OpenAI’s GPT-6.1 Sol and GPT-6 Luna, Anthropic’s Claude Haiku 5.5 and Claude Opus 5.5, and Google’s Gemini 3.5 Flash-Lite.

Datasets

The study did not use a standard pre-existing dataset. Instead, it generated a new behavioral dataset consisting of 64,302 decisions made by LLM agents within the custom-built resource-sharing simulator. This environment is designed to mimic real-world infrastructure challenges like stale load information and limited service windows.

Key Insights and Ideas

  • Information isn't enough: Simply telling agents that a server has recovered (or showing them proof) often makes performance worse at a population level. Agents flock to the newly available resource, oversubscribing it and leaving the secondary server (B) completely idle.
  • The Stampede Effect: As seen in Figure 4, when a success record is shared, agents shift from 0% choosing A to nearly 100%. This concentration causes a loss of capacity because the agents fail to coordinate who should stay on the smaller, reliable server B.
Figure 4: A shared success record improves individual discovery but causes a stampede that leaves Server B idle.
  • Algorithmic Monoculture: Because LLMs are often deterministic or follow similar probabilistic patterns, giving them the same input results in the same output. In a resource-sharing context, this lack of diversity is a bug, not a feature.
  • Implicit Coordination via Noise: The most effective way to improve efficiency was giving agents distinct private numbers. Even without instructions on how to use the numbers, advanced models like Sol and Gemini used them to diversify their choices, effectively "splitting the difference" and filling both servers.

Key Results

  • Discovery: Providing a short description of possible recovery (KT) sharply reduced missed deadlines. In Luna, missed deadlines dropped from ~60% to ~7% when request counts were visible.
  • The Concentration Cost: In the "shared success" tests, all five models left capacity unused. For example, Sol agents unanimously chose A, completing only 6 jobs and leaving 2 units of capacity on B wasted.
  • Symmetry Breaking: Giving agents distinct numbers (Condition I) raised completions in Sol from 6.0 to 7.25 out of 8 possible jobs.
  • Grouping Matters: The benefit was not just having a number in the prompt, but having different numbers within the group. When a group shared the same number (Condition C), they remained synchronized and inefficient.
Figure 6: Distinct private numbers (Condition I) help agents spread out and complete more jobs compared to identical numbers (Condition C).

Limitations and Future Work

  • Model Variance: Not all models "understood" the private number signal. While Sol and Gemini improved, Claude Haiku and Luna remained largely unanimous in their choices regardless of the numbers provided.
  • Selfishness: Agents were instructed only to complete their own jobs, not to optimize for the group. The authors suggest that future work should explore "coordinating AIs" that act as central schedulers with a global view of the group's needs.
  • Simplicity of Environment: The test only involved two servers and one recovery event. Real-world systems are far more dynamic, with fluctuating capacities and thousands of agents.

Summary updated

Abstract

When AI agents share limited resources, such as computing services or public infrastructure, their individual choices affect how much work they complete as a group. Can sharing evidence of what is available let a group make full use of its resources after capacity changes? We tested this with language-model agents choosing between two simulated servers. One server recovered while its status report still listed it as unavailable. Each server had limited capacity per round, so rejected jobs required another attempt before the deadline. Telling agents that recovery was possible reduced missed deadlines. In controlled tests with Anthropic's Claude Haiku 5.5, no agent tried the recovered server after a rule reminder, yet all chose it when shown a successful request record. After sharing, all five models tested left capacity unused. Keeping the success record fixed, we gave agents numbers without instructions, using each value equally often but giving group members either identical or distinct numbers. With OpenAI's GPT-6.1 Sol or Google's Gemini 3.5 Flash-Lite, distinct numbers spread simultaneous requests across servers and increased job completions even though overall choice frequencies barely changed. Telling AI agents which resources are available leaves a second design challenge: helping them spread demand across those resources.
BibTeX
@misc{ezaki2026groups,
  author = {Takahiro Ezaki and Naoto Imura and Katsuhiro Nishinari},
  title = {{Groups of AI agents underuse resources despite shared evidence of increased capacity}},
  year = {2026},
  eprint = {2611.22632},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2611.22632},
}
Figure