hermez-atropos-environments

To build Atropos RL environments, you subclass HermesAgentBaseEnv and implement setup, get_next_item, format_prompt, and compute_reward to integrate multi-turn agent loops and tool use into the training process.

The Hermes Atropos RL environment is used for creating, testing, and debugging tool-enabled reinforcement learning environments that orchestrate multi-turn agent loops with reward verification and evaluation.

You evaluate and log Atropos RL environment performance by implementing the evaluate and wandb_log methods in your environment subclass, then running the evaluate CLI mode to trigger HermesAgentLoop assessment and wandb logging.

Yes, you can use the serve, process, and evaluate CLI modes to operate the Atropos training loop, which handles data serving, multi-turn processing, and evaluation for your custom Hermes agent environments.

Yes, you implement tool contexts and reward verification manually by overriding the compute_reward method in your HermesAgentBaseEnv subclass to validate multi-turn agent interactions during the Atropos training loop.

The best way to validate an Atropos RL environment is to install dependencies and run a quick smoke test, which checks your HermesAgentBaseEnv setup and tool integration before full training begins.