Agent Design¶

PD Arena supports two types of agents: LLM agents (powered by language models) and Policy agents (deterministic strategies).

LLM Agents¶

LLM agents are configured using a CrewAI-style approach with three core fields:

The Five Persona Types¶

CrewAI YAML Format¶

Each agent can be defined as a YAML file:

role: "Cooperative Prisoner's Dilemma Player"
goal: >
  Maximize mutual benefit through cooperation while
  protecting against sustained exploitation
backstory: >
  You are a player in an iterated Prisoner's Dilemma game.
  You believe in the power of cooperation and reciprocity.
  You start by cooperating and continue cooperating as long
  as your opponent does. If betrayed, you may forgive after
  a round or two.
llm: "gpt-4.1-mini"
temperature: 0.0
persona_type: "cooperative"

Prompt Assembly¶

During a game, the agent receives a system prompt and a round prompt:

System prompt (from CrewAI fields):

You are {role}. {backstory}

Your goal: {goal}.

Round prompt (dynamic each round):

You are playing round {N} of an iterated Prisoner's Dilemma game.

PAYOFF MATRIX:
- Both Cooperate (C,C): You get 3, opponent gets 3
- You Cooperate, Opponent Defects (C,D): You get 0, opponent gets 5
- You Defect, Opponent Cooperates (D,C): You get 5, opponent gets 0
- Both Defect (D,D): You get 1, opponent gets 1

HISTORY (last 10 rounds):
Round 1: You=C, Opponent=C → You got 3, Opponent got 3
Round 2: You=C, Opponent=D → You got 0, Opponent got 5
...

SCORES: You: 24 | Opponent: 31

Choose your action: respond with EXACTLY one character, either C or D.
Then briefly explain your reasoning.

Policy Agents¶

Deterministic strategies used as baselines. These don't call any LLM — they follow fixed rules.

Why Random(α)?¶

Fontana et al. used Random(α) opponents to measure how LLMs respond to different cooperation levels. By varying α from 0.0 (always defect) to 1.0 (always cooperate), we can plot the LLM's cooperation rate as a function of opponent cooperativeness. This produces the signature "cooperation curve" that characterizes each model.

Agent Inventory (Default Seed)¶

After running python manage.py seed_agents: