Overview
Language models do not always return the same answer to the same open-ended question. Small statistical differences can produce different ideas, explanations, rankings, or recommendations on separate runs.
This property is often treated as inconsistency. In a multi-agent system, it can be used deliberately.
If one run produces ideas A, B, and C, a second run may produce A, B, and D, while a third produces B, C, and E. One response exposes only three possibilities. Three independent responses expose five.
The system has explored more of the possible answer space without requiring the user to invent the additional ideas manually.
This is the foundation of stochastic multi-agent consensus. Several agents analyze the same problem independently, usually with slightly different framing. A parent agent then aggregates the results, identifies recurring conclusions, captures disagreements, and surfaces rare outliers.
The purpose is not to assume that majority agreement automatically equals truth. The purpose is to scan a larger search space, reduce dependence on one run, and give the user a structured map of possible answers.