AWorld

AWorld (Agent World) from inclusionAI is a framework for building agents and the environments they operate in. The README presents four parts of what it calls the recipe for agentic learning: environment access, agent construction, experience retrieval (turning runs into training samples) and model training, with the idea that agents can use the same components to improve themselves. Install is from source: clone the repo and `pip install -e .`, with Python 3.11 or newer.

An agent is described as a model plus tools. You need a model endpoint and an environment, such as the project's hosted MCP environments or your own MCP toolchain, and the README's example builds a search agent with `Agent(name=..., system_prompt=..., mcp_config=...)` and runs it with `Runners.sync_run`. LLM credentials are set with `LLM_MODEL_NAME`, `LLM_API_KEY` and `LLM_BASE_URL`. It also covers workflows, MCP-enabled agents and multi-agent systems, with a pre-built DeepResearch team in an online playground.

For training, the runtime records full task trajectories (every LLM call, action and reward) and a `step()` mode for single-step inspection, so you can synthesize samples and inject intermediate rewards. On environments, the README lists a hosted default setup with a limited free tier for tools that cost money, while bring-your-own keys and self-hosted Docker images are marked coming soon. It is MIT licensed and was last pushed on 2026-10-12. This review is based on the README and repository data; we have not run it ourselves.