Introduction

We organize it around four parts:

• src/gym_anything/: the core responsible for creating, managing and running environments.
• benchmarks/cua_world/: benchmark environments, tasks, and support files
• agents/: reference agent implementations and evaluation loops
• extras/research/: research methods built on top of the library (env creation, task generation)

The core idea is simple:

1. Define an environment with a config file.
2. Start it.
3. Interact with it through a simple environment API.
4. Optionally attach a task and an automatic checker.
5. Use agents or your own control loop on top.

Who This Is For

We wrote these docs for three kinds of usage:

• people who want to install the framework and try a few environments
• people who want to modify environments, tasks, or agent loops
• people who want to contribute to the repository itself

The structure reflects that progression. The Start section gets you running, the Core / Benchmarks / Agents sections explain the main parts of the system, and Contributing plus Reference go deeper.

Mental Model

The framework should be understood in this order:

Start Here