zedclaw-atropos-environments

To build a multi-turn tool-calling agent environment for Atropos training, you implement only the required environment hooks like setup, get_next_item, format_prompt, and compute_reward while the base env handles orchestration and tool resolution.

Reward function verification with ToolContext scores rollouts by parsing AgentResult.messages for final responses and tool usage, optionally validating outcomes inside the sandbox to ensure correct reinforcement learning signals.

Yes, you can use OpenRouter or vLLM as inference providers for Atropos agentic benchmarking by selecting the correct inference provider flags when running serve, process, or evaluate CLI workflows.

The serve, process, and evaluate CLI modes are available for generating training data and running agentic benchmarking with tools, providing end-to-end workflows for reinforcement learning evaluation procedures.

No, you do not need to manually parse tool calls when implementing a ZedClawAgentBaseEnv subclass because the base environment automatically handles multi-turn orchestration and tool resolution for the agent loop.

To log evaluation metrics for reinforcement learning environments using wandb, implement the wandb_log hook within your ZedClaw environment class to track rollout scores and tool usage during the evaluate CLI mode.