Snowflake-Labs/agent-world-model

Agent World Model: Infinity Synthetic Environments for Agentic Reinforcement Learning

What it solves

Agent World Model (AWM) addresses the scarcity of diverse, executable environments for training and evaluating agents that use tools. It provides a way to automatically generate thousands of unique, SQL-backed synthetic environments that agents can interact with in real-time to improve their multi-turn tool-use capabilities through reinforcement learning.

How it works

AWM uses a multi-step synthesis pipeline to create these environments:

  1. Scenario Generation: Creates unique scenario descriptions from a small seed set.
  2. Task Generation: Defines specific user tasks for each scenario.
  3. Database Synthesis: Generates the SQL database schema and populates it with sample data to support the tasks.
  4. Interface Synthesis: Creates API specifications and the corresponding Python code to expose the environment via a unified Model Context Protocol (MCP) interface.
  5. Verification Synthesis: Generates code-based or LLM-augmented judges to verify if the agent successfully completed the task.

Who it’s for

This project is designed for AI researchers and developers working on agentic reinforcement learning, tool-use evaluation, and the development of LLMs that need to operate in complex, database-driven environments.

Highlights

  • Massive Scale: Synthesizes 1,000 executable, SQL-backed environments.
  • Unified Interface: Uses the Model Context Protocol (MCP) for consistent tool interaction.
  • RL-Ready: Integrated with meta-pytorch/OpenEnv to support large-scale agentic RL training.
  • Automated Verification: Includes both purely code-based and LLM-as-a-Judge verification methods.

Related

  • Project
    liruiw/GenSim

    GenSim is an LLM-powered pipeline that automatically generates robotic simulation tasks and environments to augment training data for multitask robotic policies.

  • Project
    QwenLM/Qwen-AgentWorld

    A native language world model that simulates agentic environments across seven domains to provide scalable and controllable training and evaluation for AI agents.

  • Project
    InternLM/InternBootcamp

    A scalable synthetic environment framework for training and evaluating LLM agents through executable, interactive, and verifiable tasks.

  • Project
    AgentR1/Agent-R1

    A modular framework for Agentic Reinforcement Learning that trains multi-step LLM agents by treating each interaction turn as a step-level MDP transition.

  • Dispatch
    Qwen-AgentWorld release: language world model for seven domains and its impact on general agents

    Qwen releases Qwen‑AgentWorld, a language world model that simulates seven agent environments and improves general agents via controllable simulation and unified next‑state prediction.