CodexPlay

A chaotic little playground for AI agent experiments.

This repo starts with Agent Arena, a toy simulation where agents battle over decisions, earn Zenkai-style upgrades from failure, and get judged by a deterministic governor before anything is allowed to win.

What is this?

CodexPlay is meant to be a sandbox for:

• agent combat mechanics
• deterministic judging
• visibility logs
• skill evolution experiments
• silly stick-figure / Tournament of Power style visualizations later

Run the arena

python -m codexplay.arena

Or install in editable mode first:

pip install -e .
python -m codexplay.arena

Current prototype

The first version includes four agents:

• GokuBuilder — creative builder, high novelty
• VegetaVerifier — ruthless critic, high rigor
• BulmaOptimizer — practical optimizer, high efficiency
• WhisGovernor — deterministic judge, keeps the Saiyans from deleting production

Each round gives the agents a challenge. Agents generate a proposal profile, clash, receive scores, and evolve slightly based on the outcome.

Example output

=== CODEXPLAY: AGENT ARENA ===
Challenge: Patch failing tests without breaking existing behavior.

GokuBuilder launches Creative Leap!
VegetaVerifier counters with Ruthless Verification!
BulmaOptimizer deploys Practical Refactor!

WhisGovernor evaluates proposals...
Winner: VegetaVerifier
Battle lesson stored: rigor beat novelty on safety-critical task.

Design rule

No real-world autonomy. No self-propagation. No uncontrolled external access.

The arena is for bounded simulations, agent evaluation, and visible decision-making.