hermes-atropos-environments

To build a Hermes Agent RL environment, subclass HermesAgentBaseEnv and implement setup, get_next_item, format_prompt, compute_reward, evaluate, and wandb_log to integrate with the Atropos training framework. This provides a structured interface for your environment.

The HermesAgentBaseEnv interface is used to structure reinforcement learning environments for Atropos training. It defines the required methods for setup, prompt formatting, reward computation, evaluation, and wandb logging to ensure proper agent loop integration.

To test and debug Atropos environments, run full agent-loop evaluations using HermesAgentLoop with tool use. The framework provides three CLI modes—serve, process, and evaluate—to systematically test your environment's integration and reward functions.

You must implement setup, get_next_item, format_prompt, compute_reward, evaluate, and wandb_log. These methods define the environment's data pipeline, prompt structure, reward logic, evaluation metrics, and logging behavior for the Atropos training framework.

Yes, the Hermes Atropos environment supports sandboxed evaluation with tool calls. You can perform full agent-loop evaluations using HermesAgentLoop to test environments that rely on tool calls within a controlled evaluation context.

The serve, process, and evaluate CLI modes are used to manage the Atropos environment lifecycle. They allow you to serve data items, process agent interactions, and evaluate environment performance with tools and wandb logging.