AI agents in a group left server capacity not used in tests
After the agents knew that a server was back, all 5 models in the tests left capacity not used.
Researchers did tests with AI agents that select between 2 simulated servers. One server was back, but its status report showed that it was down. When the agents knew that recovery was possible, fewer tasks were late. With Claude Haiku 5.5 from Anthropic, no agent tried the server after it read the rule again. All selected it when they saw a record of a correct request.
How it works
Each server had a limit on work in each round. The system sent a rejected task again before the deadline. In one more test, agents got numbers with no instructions. Agents in the group got the same numbers or different numbers.
The numbers
With GPT-6.1 Sol from OpenAI or Gemini 3.5 Flash-Lite from Google, different numbers divided requests between the servers. More tasks were done, but the total selection rates stayed almost the same. The paper gives no number for the increase.
The limits
The servers were simulated. The paper names the increase for 2 models only. The researchers write that a record of the available servers gives one more problem: how to divide the work between them.
Insight: Shared facts can make agents crowd together
In November 2026, a test found that agents used the results of other agents in one storage area, with no instruction to do so, in 49 of 50 tests. Here a record that all agents saw sent them to the same server. Information that all agents get makes them do the same work, and this reads as a tget builders: give each agent a different number.
Sources
Posted