Agent World Model (AWM) provides a pipeline for synthesizing a large-scale collection of 1,000 executable, SQL database-backed environments tailored for agentic reinforcement learning. It addresses the need for diverse, realistic, and verifiable training grounds for multi-turn tool-use agents, benefiting researchers and engineers in AI development.

How It Works

AWM employs a multi-stage synthesis process: starting from high-level scenarios, it generates user tasks, synthesizes SQLite databases with schema and sample data, creates a Python interface layer via FastAPI and MCP, and finally generates verification code for reward signals. This approach enables the creation of fully synthetic, executable environments with a unified interface and integrated reward mechanisms, facilitating large-scale agent training.

Quick Start & Requirements

• Primary Install: Use uv sync to set up the Python environment.
• Prerequisites: LLM API credentials (e.g., OpenAI, Azure OpenAI) are required for synthesis. Environment variables like AWM_SYN_LLM_PROVIDER, OPENAI_API_KEY, AZURE_ENDPOINT_URL must be configured. vLLM is recommended for serving models for agent demos.
• Pre-synthesized Environments: Download the 1,000 environments from Hugging Face: hf download Snowflake/AgentWorldModel-1K --repo-type dataset --local-dir ./outputs/.
• Links:
  • Paper: https://arxiv.org/abs/2602.10090
  • Code: https://github.com/Snowflake-Labs/agent-world-world-model
  • Environments: https://huggingface.co/Snowflake/AgentWorldModel-1K
  • Models: https://huggingface.co/Snowflake/Arctic-AWM-4B, https://huggingface.co/Snowflake/Arctic-AWM-8B, https://huggingface.co/Snowflake/Arctic-AWM-14B

Highlighted Details

• 1,000 executable, SQL database-backed tool-use environments generated synthetically.
• Unified MCP interface for consistent action and observation spaces across environments.
• Comprehensive synthesis pipeline covering scenario, task, database, interface, and verification code generation.
• Associated Arctic-AWM LLM models (4B, 8B, 14B) provided for agent demonstrations.

Maintenance & Community

No specific community channels (e.g., Discord, Slack), roadmap links, or details on project maintainers beyond the listed authors are provided in the README.

Licensing & Compatibility

The README does not explicitly state the project's license. This omission requires further investigation for commercial use or integration into closed-source projects.

Limitations & Caveats

The synthesis process relies heavily on LLMs, potentially introducing biases or limitations inherent to the models used. As synthetic environments, they may not fully capture the complexity or edge cases of real-world scenarios. The project requires LLM API access and potentially significant computational resources for running agent demos with large models. No license information is provided.