I was working on a local LLM ReAct agent that can play the prisoner's dilemma. The "thought" process was fascinating to watch. Not to anthropomorphize though.
Can you ask the agents to be lenient, or stringent or forgetful or forgiving ? And then have one group who participate in decision play against other group. The groups will have distribution of the properties I laid out before or other interesting properties.
I have been desiring to know what would human like features do to prisoners dilemma strategies after watching veritasium video.
What Game Theory Reveals About Life, The Universe, and Everything
