What problem does it solve?
It helps you build and troubleshoot Zeus Agent RL environments for the Atropos training framework, especially when you need a full multi-turn Hermes-style agent loop with tool calling and reward verification.
Core Features & Use Cases
- HermesAgentBaseEnv integration: Guides you on implementing only the required environment methods while the base env orchestrates the agent loop and tool resolution.
- Reward computation patterns: Shows how to score rollouts via LLM judgment, sandbox/terminal verification, or multi-signal reward composition.
- Tool-aware evaluation loop: Ensures evaluation uses the HermesAgentLoop (not single-turn chat completion) and supports deterministic evaluation settings.
- End-to-end CLI modes: Covers serve, process (offline JSONL generation), and evaluate (periodic/standalone evaluation) workflows including logging (wandb) guidance.
Quick Start
Ask what inference provider you want to use, then implement your environment in environments/your_env.py by wiring up setup(), get_next_item(), format_prompt(), compute_reward(), evaluate(), and wandb_log() following the HermesAgentBaseEnv architecture.