Arena System

Adversarial self-improvement for AI agents. Reduces hallucinations through Agent vs Anti-Agent debate loops.

Adversarial self-improvement framework for AI agents.

What it does

Give one agent two personas:

• Agent - Does the work, writes reports
• Anti-Agent - Questions everything, writes counter-reports

They take turns critiquing each other until you stop them.

Why use it

AI agents are overconfident. They hallucinate. Arena forces them to question their own outputs by arguing with themselves.

Setup

bash

./setup.sh ~/my-arena

Creates:

text

my-arena/
├── state.json
├── prompts/agent.md
├── prompts/anti-agent.md
└── outputs/

Usage

Add to HEARTBEAT.md:

1. Read state.json → whose turn?
2. Run that persona
3. Write to outputs/{role}/iteration_N.md
4. Switch turns, save state

Config

state.json:

json

{
  "current_turn": "agent",
  "iteration": 0,
  "topic": "my-project",
  "active": true,
  "max_iterations": 10
}

Results

Prevents premature deployments, catches bugs, forces proper validation before going live.