What problem does it solve? Building reinforcement learning environments that integrate with the Atropos training framework requires implementing a specific interface (setup, get_next_item, format_prompt, compute_reward, evaluate, wandb_log) and avoiding subtle pitfalls like misreading AgentResult fields or polluting training metrics during evaluation. This Skill provides the complete implementation guide, reference documentation, and testing patterns to create working Hermes Agent RL environments. ## Core Features & Use Cases - Environment Implementation Guide: Covers the HermesAgentBaseEnv interface, required methods, config classes, and the three CLI modes (serve, process, evaluate) with correct flag usage per inference provider. - Reward Function Patterns: Provides LLM-judge, binary verification via ToolContext sandbox terminal access, and multi-signal weighted scoring patterns. - Pitfall Prevention: Documents 11 common mistakes such as accessing nonexistent AgentResult fields, using chat_completion instead of HermesAgentLoop in evaluate(), and forgetting health_check=false for OpenRouter. - Use Case: You want to create a new RL environment where an agent answers research questions using tools. Follow the minimum implementation checklist, test with process mode against your chosen inference endpoint, then run evaluate mode to benchmark models. ## Quick Start Ask the agent to create a new Atropos RL environment in the hermes-agent repo implementing the HermesAgentBaseEnv interface with a reward function and evaluation loop.
