Adversarial subagents with same model? Seeking harness for different models

The author likes adversarial subagent loops where agents communicate and refine plans until consensus, but worries that using the same model for both can lead to false consensus. Asks for a good harness that supports different models (e.g., Jony Ive to Fable 5, Masayoshi Son to Kimi K3). Notes that famous people's personas are likely already encoded in LLMs.

• LLM
• 多智能体
• 对抗性子代理
• 模型不同
• 工具推荐