AWorld: The Agent Harness for Your World

"The Next Frontier for AI is Your Expertise"

Twitter Follow WeChat QR Code Discord License: MIT DeepWiki Tutorial

中文版 | Automation | Evolution | Contributing |

This is the AWorld Thesis: A powerful harness is not enough. True AI scaling is unlocked only when experts like you embed the invaluable knowledge, effectively building the gate in that wall.

AWorld is the platform designed for this singular purpose. We provide a complete, battle-tested Harness as the recipe for you, the expert, to forge your knowledge into a fleet of autonomous agents. Together, we move beyond AI's generic promise to create robust, precise applications that master your specific domain.

From Expertise to Product

See what happens when expert knowledge is encoded into reusable Skills. The creations below are orchestrated by the AWorld Agent, demonstrating our core scaling law: as the community contributes more expertise, the entire ecosystem becomes more powerful.

From one-prompt video generation to deep-search workflows, each example turns specialized know-how into repeatable production capability.

This is what's possible today. Imagine what we'll build with your expertise.

Your Journey with AWorld-CLI

The journey from an idea to an evolved, autonomous agent begins at your fingertips.

Install and Activate

Install once, configure globally, and run anywhere.

Install AWorld-CLI

git clone https://github.com/inclusionAI/AWorld && cd AWorld

conda create -n aworld_env python=3.11 -y && conda activate aworld_env 

pip install -e . && cd aworld-cli && pip install -e .

Config & Launch

cd your working directory

aworld-cli --config

Once configured, simply type aworld-cli in your terminal to start your journey.

Alternatively, you can configure by creating a .env file in your working directory with your model and API settings. See AWorld CLI Configuration for the core variables.

Automate Creation with AWorld-CLI

This multi-agent system works in concert to turn your ideas into reality:

The Evolution Loop: Build -> Evaluate -> Evolve

Imagine you ask: "Help me create an English word learning mini-app with a UI quality score above 0.9."

  • The Developer Builds: The Developer analyzes requirements and writes code (e.g., HTML) using CAST.
  • The Evaluator Judges: The Evaluator inspects the output using our verified Skill.
  • The Loop Refines: If the score is below target (e.g., 0.9), AWorld instructs the Developer to fix specific issues identified by the Evaluator. This loop continues until your criteria are met.

📹 See the Self-Evolution Loop in Action

No Evaluation, No Evolution

CAST: Conquering Code Complexity

  • Hierarchical Navigation: Instantly understand code structure and purpose without getting lost in implementation details.
  • Nearly Infinite Context: Intelligently compresses code to feed the agent only relevant information, breaking the context window limitation.
  • Surgical Code Modification: Perform precise changes with full dependency awareness, avoiding the clumsy errors of "blind" text replacement.

Your Expertise as the Evaluator

A Proven Harness: Benchmark Excellence

Agent Benchmarking

Data Synthesis

  1. FunReason-MT Technical Report: Overcoming the Complexity Barrier in Multi-Turn Function Calling arxiv, 2025. paper, code, model, dataset

    Zengzhuang Xu, Bingguang Hao, Zechuan Wang, Yuntao Wen, Maolin Wang, etc.

  2. From Failure to Mastery: Generating Hard Samples for Tool-use Agents arxiv, 2026. paper, code, model, dataset

    Bingguang Hao, Zengzhuang Xu, Yuntao Wen, Xinyi Xu, Yang Liu, etc.

Model Training

  1. AWorld: Orchestrating the Training Recipe for Agentic AI. arxiv, 2025. paper, code, model

    Chengyue Yu, Siyuan Lu, Chenyi Zhuang, Dong Wang, Qintong Wu, etc.

  2. FunReason: Enhancing Large Language Models' Function Calling via Self-Refinement Multiscale Loss and Automated Data Refinement. arxiv, 2025. paper, model

    Bingguang Hao, Maolin Wang, Zengzhuang Xu, Cunyin Peng, etc.

  3. Exploring Superior Function Calls via Reinforcement Learning. arxiv, 2025. paper, code

    Bingguang Hao, Maolin Wang, Zengzhuang Xu, Yicheng Chen, etc.

  4. RAG-R1 : Incentivize the Search and Reasoning Capabilities of LLMs through Multi-query Parallelism. arxiv, 2025. paper, code, model

    Zhiwen Tan, Jiaming Huang, Qintong Wu, Hongxuan Zhang, Chenyi Zhuang, Jinjie Gu

  5. V2P: From Background Suppression to Center Peaking for Robust GUI Grounding Task. arxiv, 2025. paper, code

    Jikai Chen, Long Chen, Dong Wang, Leilei Gan, Chenyi Zhuang, Jinjie Gu

  6. Don’t Just Fine-tune the Agent, Tune the Environment arxiv, 2025. paper

    Siyuan Lu, Zechuan Wang, Hongxuan Zhang, Qintong Wu, Leilei Gan, Chenyi Zhuang, etc.

Meta Learning

  1. Profile-Aware Maneuvering: A Dynamic Multi-Agent System for Robust GAIA Problem Solving by AWorld. arxiv, 2025. paper, code

    Zhitian Xie, Qintong Wu, Chengyue Yu, Chenyi Zhuang, Jinjie Gu

  2. Recon-Act: A Self-Evolving Multi-Agent Browser-Use System via Web Reconnaissance, Tool Generation, and Task Execution. arxiv, 2025. paper, code

    Kaiwen He, Zhiwei Wang, Chenyi Zhuang, Jinjie Gu

Advanced Features

Background Subagent Execution (v2.0)

Non-blocking orchestration: Spawn subagents and continue working while they execute in the background.

from aworld.core.common import ActionModel
from aworld.core.tool.builtin.spawn_subagent_tool import SpawnSubagentTool

# Initialize tool
spawn_tool = SpawnSubagentTool(subagent_manager=your_manager, conf=config)

# 1. Start background task (returns immediately)
action = ActionModel(
    action_name='spawn_background',
    params={
        'name': 'deep_researcher',
        'directive': 'Comprehensive research on quantum computing',
        'task_id': 'research_quantum'
    }
)
obs, reward, _, _, info = await spawn_tool.do_step([action])
task_id = info['task_id']

# 2. Orchestrator continues other work (parallel execution!)
# ... analyze existing data, plan next steps, spawn more tasks ...

# 3. Check status (non-blocking)
check_action = ActionModel(
    action_name='check_task',
    params={'task_id': task_id, 'include_result': False}
)
obs, reward, _, _, info = await spawn_tool.do_step([check_action])
print(f"Status: {info['status']}, Elapsed: {info['elapsed']:.2f}s")

# 4. Wait for completion when ready
wait_action = ActionModel(
    action_name='wait_task',
    params={'task_ids': task_id, 'timeout': 300}
)
await spawn_tool.do_step([wait_action])

# 5. Get result
check_action = ActionModel(
    action_name='check_task',
    params={'task_id': task_id, 'include_result': True}
)
obs, reward, _, _, info = await spawn_tool.do_step([check_action])
result = info['result']

Performance Benefit: Orchestrator work and subagent execution overlap, achieving max(T_orchestrator, T_subagent) instead of T_orchestrator + T_subagent. Benchmark shows 3x faster execution for typical workloads.

Documentation: See docs/Agents/Parallel Subagents.md

Contributing

We warmly welcome developers, researchers, and domain experts to join us. Whether you're enhancing the framework or contributing a Skill from your field of expertise, your work is valuable.

Historical planning notes live under aworld-cli/internal/docs/superpowers/.

For academic citations or wish to contact us, please use the following BibTeX entry:

@misc{yu2025aworldorchestratingtrainingrecipe,
      title={AWorld: Orchestrating the Training Recipe for Agentic AI}, 
      author={Chengyue Yu and Siyuan Lu and Chenyi Zhuang and Dong Wang and Qintong Wu and Zongyue Li and Runsheng Gan and Chunfeng Wang and Siqi Hou and Gaochi Huang and Wenlong Yan and Lifeng Hong and Aohui Xue and Yanfeng Wang and Jinjie Gu and David Tsai and Tao Lin},
      year={2025},
      eprint={2508.20404},
      archivePrefix={arXiv},
      primaryClass={cs.AI},
      url={https://arxiv.org/abs/2508.20404}, 
}