Agent-World: Scaling Real-World Environment Synthesis for Evolving General Agent Intelligence
Guanting Dong, Junting Lu, Junjie Huang, Wanjun Zhong, Longxiang Liu, Shijue Huang, Zhenyu Li, Yang Zhao, Xiaoshuai Song, Xiaoxi Li, Jiajie Jin, Yutao Zhu, Hanbin Wang, Fangyu Lei, Qinyu Luo, Mingyang Chen, Zehui Chen, Jiazhan Feng, Ji-Rong Wen, Zhicheng Dou
2026-04-21
Summary
This paper introduces Agent-World, a new system designed to train artificial intelligence agents to be more generally intelligent and capable of using tools in the real world.
What's the problem?
Currently, training AI agents to be truly helpful and adaptable is difficult because we lack realistic environments for them to learn in and good ways to help them continuously improve over time. Existing training methods struggle to create agents that can handle a wide variety of tasks and situations, and it's hard to keep them learning and getting better as the world changes.
What's the solution?
The researchers created Agent-World, which has two key parts. First, it automatically finds and creates tasks for the agents to practice, drawing from many different real-world scenarios and tools. Second, it constantly monitors how the agents are doing and creates new, more challenging tasks to push them to learn and improve, essentially creating a self-improving learning loop where both the agent and the training environment evolve together.
Why it matters?
This work is important because it shows a promising way to build AI agents that are more versatile and capable of handling complex tasks in the real world. The Agent-World system significantly outperformed other AI models on a variety of tests, and the insights gained from it can help us develop even more intelligent and adaptable AI in the future.
Abstract
Large language models are increasingly expected to serve as general-purpose agents that interact with external, stateful tool environments. The Model Context Protocol (MCP) and broader agent skills offer a unified interface for connecting agents with scalable real-world services, but training robust agents remains limited by the lack of realistic environments and principled mechanisms for life-long learning. In this paper, we present Agent-World, a self-evolving training arena for advancing general agent intelligence through scalable environments. Agent-World has two main components: (1) Agentic Environment-Task Discovery, which autonomously explores topic-aligned databases and executable tool ecosystems from thousands of real-world environment themes and synthesizes verifiable tasks with controllable difficulty; and (2) Continuous Self-Evolving Agent Training, which combines multi-environment reinforcement learning with a self-evolving agent arena that automatically identifies capability gaps through dynamic task synthesis and drives targeted learning, enabling the co-evolution of agent policies and environments. Across 23 challenging agent benchmarks, Agent-World-8B and 14B consistently outperforms strong proprietary models and environment scaling baselines. Further analyses reveal scaling trends in relation to environment diversity and self-evolution rounds, offering insights for building general agent intelligence.