AgentWorldModel-1K

Agent World Model: Infinity Synthetic Environments for Agentic Reinforcement Learning

Zhaoyang Wang¹, Canwen Xu², Boyi Liu², Yite Wang², Siwei Han¹,
Zhewei Yao², Huaxiu Yao¹, Yuxiong He²

¹UNC-Chapel Hill   ²Snowflake AI Research  

Overview

AgentWorldModel-1K contains 1,000 fully synthetic, executable, SQL database-backed tool-use environments exposed via a unified MCP (Model Context Protocol) interface, designed for large-scale multi-turn agentic reinforcement learning.

Each environment is synthesized through the Agent World Model (AWM) pipeline:

1. Scenario — A high-level description (e.g., "an online shopping platform")
2. Tasks — 10 user tasks per scenario that serve as functional requirements
3. Database — SQLite database schema and sample data as the state backend
4. Interface — Python interface layer (FastAPI + MCP) as the action/observation space
5. Verification — Verification code that inspects database state changes for reward signals

For the full synthesis pipeline, please visit https://github.com/Snowflake-Labs/agent-world-model.

Resources

Related resources are also available, please check:

Dataset Files

Citation

If you find this resource useful, please kindly cite:

@article{wang2026agentworldmodelinfinity,
      title={Agent World Model: Infinity Synthetic Environments for Agentic Reinforcement Learning}, 
      author={Zhaoyang Wang and Canwen Xu and Boyi Liu and Yite Wang and Siwei Han and Zhewei Yao and Huaxiu Yao and Yuxiong He},
      year={2026},
      eprint={2602.10090},
      archivePrefix={arXiv},
      primaryClass={cs.AI},
      url={https://arxiv.org/abs/2602.10090}, 
}