Skip to content

WRL-Dragon

WRL-Dragon automates reinforcement learning policy development using a hierarchical multi-agent interaction system.

Code Demo Video Cerebral Valley

Core trick

A CEO agent delegates to Coder agents that generate RL policy code, and QA agents run Gym rollouts to feed scores into the next round.

Tags

agentsdev tools

Description

WRL-Dragon automates reinforcement learning policy development using a hierarchical multi-agent interaction system. Instead of having to manually write and tune RL policies, a CEO agent analyzes environments, Coder agents generate policy code using LLMs, and QA agents run rollouts to evaluate performance. Results feed back into subsequent round, creating an autonomous improve loop across multiple Gym environments simultaneously. The system solves the tedious cycle of write-test-iterate in RL research by letting AI agents handle the entire pipeline from strategy to code generation to evaluation, all while a real-time dashboard lets you watch the process unfold. Beyond this orchestration loop, the system meta-trains smaller child agents directly on RL tasks, combining LLM-driven code generation with traditional RL training to produce increasingly capable policies!
farhan-navas/wrl-dragon★ 00 forksPythonCreated Mar 7, 2026 · last push Mar 8, 2026

README excerpt

# wrl-dragon

RL orchestrator that uses Claude Code to coordinate sub-agents running OpenEnv environments, with a visual observability platform.

## Overview

### Phase 1 — Claude Code Orchestrator (current)

Claude Code acts as a "CEO" orchestrator that:
- Generates task-specific code, reward functions, and environment configs
- Spawns sub-agents that each run their own OpenEnv environment
- Collects logs and renders videos of agent rollouts

**Observability platform** — a web app (ChatDev / PixelAgents style) that visualizes the hierarchy:
- Top-level view: CEO orchestrator, coder agents, worker agents
- Click on any worker to see visualized rollouts across rounds
- Real-time logs, reward curves, and video playback

```
┌─────────────────────────────────┐
│  Claude Code (CEO Orchestrator) │
│  - generates reward functions   │
│  - assigns tasks to sub-agents  │
│  - aggregates results           │
└──────────┬──────────────────────┘
           │
     ┌─────┴─────┐
     ▼           ▼
┌─────────┐ ┌─────────┐
│ Coder 1 │ │ Coder 2 │  ← generate env code, reward fns
└────┬────┘ └────┬────┘
     ▼           ▼
┌─────────┐ ┌─────────┐
│Worker 1 │ │Worker 2 │  ← run OpenEnv tasks, produce rollouts
└─────────┘ └─────────┘
     │           │
     ▼           ▼
  [logs, videos, reward curves]
         │
         ▼
  ┌──────────────┐
  │  Dashboard   │
  │  (web app)   │
  └──────────────┘
```

### Phase 2 — RL Meta-Training (current)

Replace Claude Code with GRPO (Group Relative Polic

Team