⚔️ AgentArena

Gamified Agentic Coding — Vibe Coding on Steroids

Turn your coding tasks into RPG battles where AI agents fight kernel-level enemies.

AgentArena transforms the agentic coding experience into a gamified terminal RPG. Your AI agents — CodeSmith, BugSlayer, Architect, TestKnight, DeployMage, and DocScribe — battle against task enemies with HP, defense, and boss mechanics. Every coding subtask is a quest, every module is an enemy to defeat, and token efficiency determines your victory rank.

# Install
pip install -e .

# Run the demo battle
python run_demo.py

# Or use the CLI
arena battle
arena party
arena version

Built on research-backed foundations:

• Predictive Coding — Agents learn via local prediction errors (Hebbian updates), not global backprop
• MaxEnt Belief Engine — Random Worlds / Maximum Entropy for rational task dispatch
• TLPlan Temporal Logic — Temporal constraints ensure correct quest ordering (no deploying before testing!)
• Octalysis Reward System — 8 core drives mapped to gamification mechanics
• InfoRM Anti-Reward-Hacking — Information bottleneck prevents gaming the token budget
• 5-Layer Context Defense — Input filtering, memory pointers, observation masking, drift detection, Merkle checkpoints

Token efficiency = winning strategy:

agentarena/
├── src/agentarena/
│   ├── models/          # Agent, Enemy, Quest, BattleState, Scratchpad
│   ├── engine/          # BattleEngine (combat loop, damage calc, XP)
│   ├── tui/             # Rich TUI (battle screen, panels, victory)
│   ├── agents/          # LangGraph agent integration (Phase 2.5)
│   ├── rag/             # Agentic RAG engine (Phase 2.5)
│   ├── config/          # Configuration and knowledge base
│   └── cli.py           # Typer CLI entry point
├── tests/
├── docs/
├── run_demo.py          # Quick demo runner
└── pyproject.toml

MIT