zedclaw-atropos-environments

What problem does it solve?

This Skill helps you build and debug ZedClaw reinforcement learning environments for Atropos so your reward signals and tool-using agent loop work correctly.

Core Features & Use Cases

• Agent-loop environment architecture: Guides you to implement only the required environment hooks while the base env handles multi-turn orchestration and tool resolution.
• Reward verification with tool context: Shows how to score rollouts using the full AgentResult (via message parsing) and optionally validate outcomes inside the sandbox with ToolContext.
• End-to-end evaluation modes: Provides the correct serve, process, and evaluate workflows to run training data generation and agentic benchmarking with tools.

Quick Start

Ask the AI to generate the minimal ZedClaw Atropos environment class that implements setup, get_next_item, format_prompt, compute_reward, evaluate, and wandb_log, then run evaluate mode with your chosen OpenAI-compatible inference configuration.