hermes-atropos-environments

Multi-turn agent loops in Atropos RL environments are built using HermesAgentBaseEnv templates that integrate tool resolution and ToolContext to handle sequential tool calling within the reinforcement learning training cycle.

Formatting prompts and computing rewards for Hermes RL training requires implementing the format_prompt and compute_reward methods within the Atropos environment to generate correct agent inputs and evaluate tool usage.

Running a full evaluation workflow for Atropos tool-calling environments involves using the evaluate CLI mode with HermesAgentLoop, applying deterministic evaluation settings, and logging results via integrated wandb tracking.

Yes, wandb logging tracks tool usage and rewards in Atropos environments through the designated wandb_log method, capturing evaluation metrics and agent loop performance data for reproducible reinforcement learning analysis.

The Hermes Atropos environments CLI supports serve, process, and evaluate modes to manage data serving, environment processing, and deterministic evaluation workflows for reinforcement learning agent loops.

Configuring multi-turn tool resolution in Hermes RL environments requires strict adherence to the HermesAgentBaseEnv structure and ToolContext integration, limiting modifications to the predefined agent loop and reward computation logic.