Agent-World Training Arena (3-min read)

Agent-World is a self-evolving training arena that collects real tool ecosystems at scale to synthesize environments and tasks, then continuously evolves agents on top of them.

The core rests on two axes.

• Agentic Environment-Task Discovery: discovers environment themes from MCP servers, tool docs, and industrial PRDs to build 2,000+ environments across 20 major categories.
• Continuous Self-Evolving Agent Training: combines multi-environment RL with diagnostic task generation to feed the agent's weaknesses back into the next training round.

The environment system is organized into a 3-level taxonomy. Under the 20 primary categories are detailed sub-classifications, ultimately linked to concrete MCP server instances, which widens coverage while systematically managing difficulty.

The database is not a synthetic DB imagined by an LLM, but is built by a deep-research agent repeatedly exploring and refining structured data from the web. Afterward, database complexification expands the content over multiple rounds.

Tools and tasks are also automatically verified.

• A tool-design agent creates candidate tools and unit tests, keeping only those that pass compilation success, test accuracy above 0.5, and a minimum of 1 tool/1 test condition, securing 19K+ tools.
• Task synthesis splits into graph-based and programmatic approaches, where the former uses a tool dependency graph and random walk, and the latter uses executable Python solutions and verification scripts.
• For graph-based tasks, an LLM writes the task and correct answer, then a ReAct agent is run 5 times to verify consistency, with difficulty adjusted via tool chain length, non-linear reasoning requirements, and tool name masking.

During the training phase, the LLM policy, tool runtime, and database state are connected in a single closed loop. Graph tasks are evaluated with a rubric-conditioned LLM-as-judge, while programmatic tasks are evaluated with executable validation scripts in a sandbox, producing structured verifiable rewards, and learning is performed with GRPO.

The results are also strong. Across 23 benchmarks, it consistently outperformed strong proprietary models and environment scaling baselines, with evaluation coverage spanning agentic tool use, advanced AI assistant, software engineering, deep research, and reasoning benchmarks. A scaling trend was also confirmed, where performance rises as environment diversity and self-evolution rounds increase.