hermes-atropos-environments

To build RL environments for Atropos training, you implement the HermesAgentBaseEnv interface, define custom reward functions, and wire the environment to support tool-enabled evaluation. Validation requires running deterministic agent loops via the three CLI modes: serve, process, and evaluate.

The HermesAgentBaseEnv interface provides the foundational scaffolding for reinforcement learning environments, orchestrating the environment loop, defining reward calculation logic, and ensuring deterministic behavior suitable for benchmarking Hermes agents during Atropos training.

You set up wandb logging by configuring the evaluation loop within your HermesAgentBaseEnv implementation. The environment supports wandb logging natively alongside its three CLI modes—serve, process, and evaluate—allowing you to track benchmarking metrics during agent training and validation.

Yes, Atropos environments are designed for deterministic benchmarking and tool-enabled evaluation. By implementing HermesAgentBaseEnv, you ensure deterministic behavior across item sequencing, setup, and logging, which is essential for generating reproducible benchmark results during agent training.

The serve, process, and evaluate CLI modes are the three operational commands for Hermes RL environments. Serve initializes the environment, process handles item sequencing and data flow, and evaluate executes the agent loop to validate performance and produce benchmarking results.

No external dependencies are required to implement the core Hermes agent loop for Atropos. The skill provides the HermesAgentBaseEnv interface and orchestration logic internally, allowing you to define reward functions and evaluation workflows without installing additional external libraries.