hermes-atropos-environments

To build a reinforcement learning environment for Atropos training, implement the required setup, item iteration, prompt formatting, and reward calculation methods using a HermesAgentBaseEnv-compatible interface for multi-turn tool-calling agent evaluation.

Reward computation fails when using non-existent fields; you must score rollouts by extracting proper AgentResult fields directly from AgentResult.messages, including tool verification via ToolContext.

You evaluate a multi-turn tool-calling agent by running an evaluate() loop using HermesAgentLoop, computing eval metrics, and logging training and eval signals to wandb for offline processing workflows.

Yes, you can use HermesAgentBaseEnv for multi-turn agent evaluation and reward verification by leveraging its agent loop orchestration and tool resolution while implementing your custom environment methods.

Running serve, process, and evaluate modes requires implementing setup, get_next_item, format_prompt, compute_reward, evaluate, and wandb_log methods to ensure full agent-loop orchestration and metric logging.

A key limitation is that reward calculation must strictly align with AgentResult.messages rather than custom fields, and proper ToolContext verification is required to ensure tool-calling rollouts are scored correctly during training.