gauss-atropos-environments

Build an RL environment for Atropos by subclassing the GaussAgentBaseEnv interface to implement multi-turn agent loops with tool calling, structured reward computation, and evaluation workflows. This provides the core interface for rapid development and benchmarking.

The best way to test and debug Gauss RL environments is using the three CLI modes: serve for running the environment, process for data generation, and evaluate for sandboxed evaluation with full agent loops. These modes enforce reproducible workflows during development.

Yes, the environment interface supports tool calling integration within the multi-turn agent loop and wandb logging for tracking metrics. Tool context is provided natively, and evaluation loops run with full agent loop capabilities for comprehensive benchmarking.

Structured reward computation uses AgentResult data to calculate rewards in Gauss environments during the evaluation phase. This ensures reproducible workflows and sandboxed evaluation by enforcing consistent data structures throughout the agent loop process.

Yes, you can run evaluation loops with full tool integration in a sandboxed environment using the evaluate CLI mode. This enforces reproducible workflows while testing the multi-turn agent loop and structured reward computation with AgentResult data.