hermes-atropos-environments

To build a custom RL environment for Atropos training, you implement the HermesAgentBaseEnv class to define your task, orchestrate the agent loop, and integrate tool-based reward evaluation. This provides the boilerplate needed for rapid prototyping.

Setting up a reinforcement learning environment with Monika agents requires Python, the monika-agent repository, and access to an LLM inference endpoint. These dependencies allow you to run the HermesAgentBaseEnv implementations.

You evaluate model performance and log rewards in an Atropos environment by implementing the compute_reward function and utilizing the evaluate CLI mode. This workflow integrates with WandB logging to track training trajectories and benchmarking metrics.

Yes, you can use the process CLI command to generate training data and trajectories for reinforcement learning. The environment also supports serve for training and evaluate for benchmarking model performance within the Monika ecosystem.

The best way to debug a custom HermesAgentBaseEnv implementation is by using the evaluate and process CLI modes to test tool-integrated reward evaluation and generate training trajectories. This isolates environment logic before full Atropos training.