Arena System
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
./setup.sh ~/my-arena
Creates:
my-arena/
├── state.json
├── prompts/agent.md
├── prompts/anti-agent.md
└── outputs/
Usage
Add to HEARTBEAT.md:
- Read
state.json→ whose turn? - Run that persona
- Write to
outputs/{role}/iteration_N.md - Switch turns, save state
Config
state.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.