File size: 24,609 Bytes

---
base_model:
- Qwen/Qwen2.5-Coder-7B-Instruct
library_name: transformers
license: mit
pipeline_tag: text-generation
---

# AWorld: Dynamic Multi-Agent System with Stable Maneuvering for Robust GAIA Problem Solving

This repository contains the model and code for **AWorld**, presented in the paper [AWorld: Dynamic Multi-Agent System with Stable Maneuvering for Robust GAIA Problem Solving](https://huggingface.co/papers/2508.09889).

Code: [https://github.com/inclusionAI/AWorld](https://github.com/inclusionAI/AWorld)

<div align="center">
  <img src="https://github.com/inclusionAI/AWorld/raw/main/readme_assets/heading_banner.png" alt="AWorld Heading Banner" width="100%">
</div>

## Abstract

The rapid advancement of large language models (LLMs) has empowered intelligent agents to leverage diverse external tools for solving complex real-world problems. However, as agents increasingly depend on multiple tools, they encounter new challenges: extended contexts from disparate sources and noisy or irrelevant tool outputs can undermine system reliability and accuracy. These challenges underscore the necessity for enhanced stability in agent-based systems. To address this, we introduce dynamic supervision and maneuvering mechanisms, constructing a robust and dynamic Multi-Agent System (MAS) architecture within the AWorld framework. In our approach, the Execution Agent invokes the Guard Agent at critical steps to verify and correct the reasoning process, effectively reducing errors arising from noise and bolstering problem-solving robustness. Extensive experiments on the GAIA test dataset reveal that our dynamic maneuvering mechanism significantly improves both the effectiveness and stability of solutions, outperforming single-agent system (SAS) and standard tool-augmented systems. As a result, our dynamic MAS system achieved first place among open-source projects on the prestigious GAIA leaderboard. These findings highlight the practical value of collaborative agent roles in developing more reliable and trustworthy intelligent systems.

## Overview

AWorld (Agent World) is the next-generation framework engineered for agent self-improvement at scale. We enable AI agents to continuously evolve by synthesizing their own knowledge and experiences. This core capability is powered by:

1.  **Multi-Agent Systems (MAS)**: Build complex, interacting agent societies using our plug-and-play protocols and robust context management.
2.  **Intelligence Beyond a Single Model**: Generates high-quality feedback and diverse synthetic training data that fuel individual agent evolution.
3.  **Cloud-Native for Diversity & Scale**: Delivers the high concurrency and scalability for training smarter agents and achieving self-improvement.

AWorld empowers you to rapidly build individual tool-using agents, orchestrate sophisticated multi-agent systems, train agents effectively, and synthesize the high-quality data required for continuous agent evolution – all converging towards autonomous self-improvement.

---
**Agentic Achievements Unboxed!** 🔥

🏅️ **[2025/08/06]** Excel in Stability with Multi-Agent System on GAIA [![][GAIA]](https://huggingface.co/spaces/gaia-benchmark/leaderboard) [![][MAS]](#) [![][Leaderboard]](#) [![][Forward]]() <br>
Achieved an average Pass@1 score of 67.89 and Pass@3 score of 83.49 across 109 tasks from the test dataset. See [more details here](./examples/gaia/README_GUARD.md). See [Technical report here](https://huggingface.co/blog/chengle/aworld-gaia).

🚀 **[2025/07/25]** Plug-&-Play Math Mastery! [![][IMO]](https://www.imo-official.org/year_info.aspx?year=2025) [![][MAS]](#) [![][Forward]]() <br>
Built a MAS solving <span style="color: #d81b60; font-weight: bold;">5/6 IMO 2025 problems</span> in hours - proving **agent orchestration** beats solo models. Peek at the genius [under the hood](examples/imo/README.md)!

<details>
<summary style="font-size: 1.2em;font-weight: bold;"> 🌏 Click to View Previous News </summary>

🧩 **[2025/07/23]** Build BFCL Runtimes Readily! [![][BFCL]](https://gorilla.cs.berkeley.edu/leaderboard.html) [![][Runtime]](#)  
New BFCL tutorial shows how to **synthesize function calls training data in minutes**. <br>
Start cooking up your own agents [right here](examples/BFCL/README.md)!

🏆 **[2025/07/07]** GAIA Soars to <span style="color: #d81b60; font-weight: bold;">77.08</span>! [![][GAIA]](https://huggingface.co/spaces/gaia-benchmark/leaderboard) [![][Cloud-Native]](#) [![][Leaderboard]](#)  
**Cloud-native runtimes** now enable distributed agent evolution. <br>
See how diverse environments evolves smarter models [training recipe](#backward-process-design).

</details>

## Quickstart

### Prerequisites
> [!TIP]
> Python>=3.11
```bash
git clone https://github.com/inclusionAI/AWorld && cd AWorld

python setup.py install
```
### Hello world examples
We introduce the concepts of `Agent` and `Runners` to help you get started quickly.
```python
import os

from aworld.agents.llm_agent import Agent
from aworld.runner import Runners

summarizer = Agent(
    name="Summary Agent", 
    system_prompt="You specialize at summarizing.",
)

result = Runners.sync_run(
    input="Tell me a succint history about the universe", 
    agent=summarizer,
)
```

In parallel, we introduce the concepts of `Swarm` to construct a team of agents.
```python
import os

from aworld.agents.llm_agent import Agent
from aworld.runner import Runners
from aworld.core.agent.swarm import Swarm

researcher = Agent(
    name="Research Agent", 
    system_prompt="You specialize at researching.",
)
summarizer = Agent(
    name="Summary Agent", 
    system_prompt="You specialize at summarizing.",
)
# Create agent team with collaborative workflow
team = Swarm(researcher, summarizer)

result = Runners.sync_run(
    input="Tell me a complete history about the universe", 
    swarm=team,
)
```

Finally, run your own agents or teams
```bash
# Set LLM credentials
export LLM_MODEL_NAME="gpt-4"
export LLM_API_KEY="your-api-key-here"
export LLM_BASE_URL="https://api.openai.com/v1"

# Run
python /path/to/agents/or/teams
```

<details>
<summary style="font-size: 1.2em;font-weight: bold;"> 🌏 Click to View Advanced Usages </summary>

### Pass AgentConfig Explicitly
```python
import os

from aworld.agents.llm_agent import Agent
from aworld.runner import Runners
from aworld.config.conf import AgentConfig
from aworld.core.agent.swarm import Swarm

gpt_conf = AgentConfig(
    llm_provider="openai",
    llm_model_name="gpt-4o",
    llm_api_key="<OPENAI_API_KEY>",
    llm_temperature=0.1,
)
openrouter_conf = AgentConfig(
    llm_provider="openai",
    llm_model_name="google/gemini-2.5-pro",
    llm_api_key="<OPENROUTER_API_KEY>",
    llm_base_url="https://openrouter.ai/api/v1"
    llm_temperature=0.1,
)

researcher = Agent(
    name="Research Agent", 
    conf=gpt_conf,
    system_prompt="You specialize at researching.",
)
summarizer = Agent(
    name="Summary Agent", 
    conf=openrouter_conf,
    system_prompt="You specialize at summarizing.",
)
# Create agent team with collaborative workflow
team = Swarm(researcher, summarizer)

result = Runners.sync_run(
    input="Tell me a complete history about the universe", 
    swarm=team,
)
```

### Agent Equipped with MCP Tools
```python
import os

from aworld.agents.llm_agent import Agent
from aworld.runner import Runners

mcp_config = {
    "mcpServers": {
        "GorillaFileSystem": {
            "type": "stdio",
            "command": "python",
            "args": ["examples/BFCL/mcp_tools/gorilla_file_system.py"],
        },
    }
}

file_sys = Agent(
    name="file_sys_agent",
    system_prompt=(
        "You are a helpful agent to use "
        "the standard file system to perform file operations."
    ),
    mcp_servers=mcp_config.get("mcpServers", []).keys(),
    mcp_config=mcp_config,
)

result = Runners.sync_run(
    input=(
        "use mcp tools in the GorillaFileSystem server "
        "to perform file operations: "
        "write the content 'AWorld' into "
        "the hello_world.py file with a new line "
        "and keep the original content of the file. "
        "Make sure the new and old "
        "content are all in the file; "
        "and display the content of the file"
    ),
    agent=file_sys,
)
```

### Agent Integrated with Memory
It is recommended to use `MemoryFactory` to initialize and access Memory instances.

```python
from aworld.memory.main import MemoryFactory
from aworld.core.memory import MemoryConfig, MemoryLLMConfig

# Simple initialization
memory = MemoryFactory.instance()

# Initialization with LLM configuration
MemoryFactory.init(
    config=MemoryConfig(
        provider="aworld",
        llm_config=MemoryLLMConfig(
            provider="openai",
            model_name=os.environ["LLM_MODEL_NAME"],
            api_key=os.environ["LLM_API_KEY"],
            base_url=os.environ["LLM_BASE_URL"]
        )
    )
)
memory = MemoryFactory.instance()
```

`MemoryConfig` allows you to integrate different embedding models and vector databases.
```python
import os

from aworld.core.memory import MemoryConfig, MemoryLLMConfig, EmbeddingsConfig, VectorDBConfig

MemoryFactory.init(
    config=MemoryConfig(
        provider="aworld",
        llm_config=MemoryLLMConfig(
            provider="openai",
            model_name=os.environ["LLM_MODEL_NAME"],
            api_key=os.environ["LLM_API_KEY"],
            base_url=os.environ["LLM_BASE_URL"]
        ),
        embedding_config=EmbeddingsConfig(
            provider="ollama", # or huggingface, openai, etc.
            base_url="http://localhost:11434",
            model_name="nomic-embed-text"
        ),
        vector_store_config=VectorDBConfig(
            provider="chroma",
            config={
                "chroma_data_path": "./chroma_db",
                "collection_name": "aworld",
            }
        )
    )
)
```

### Mutil-Agent Systems
We present a classic topology: `Leader-Executor`.
```python
"""
Leader-Executor topology:
 ┌───── plan ───┐     
exec1         exec2

Each agent communicates with a single supervisor agent, 
well recognized as Leader-Executor topology, 
also referred to as a team topology in Aworld.
"""
from aworld.agents.llm_agent import Agent
from aworld.core.agent.swarm import TeamSwarm

plan = Agent(name="plan", conf=agent_conf)
exec1 = Agent(name="exec1", conf=agent_conf)
exec2 = Agent(name="exec2", conf=agent_conf)
swarm = TeamSwarm(plan, exec1, exec2)
```
Optionally, you can use `Handsoff` mechanism to customize your own topology.
```python
from aworld.core.agent.swarm import HandoffSwarm
swarm = HandoffSwarm((plan, exec1), (plan, exec2))
```

</details>

# 🏗️ Architecture Design Principles
AWorld provides a comprehensive environment that supports a diverse array of applications, such as `Product Prototype Verification`, `Foundational Model Training`, and the design of `Multi-Agent Systems (MAS)` through meta-learning. 

This framework is engineered to be highly adaptable, enabling researchers and developers to explore and innovate across multiple domains, thereby advancing the capabilities and applications of multi-agent systems.

## Concepts & Framework
| Concepts | Description |
| :-------------------------------------- | ------------ |
| [`agent`](./aworld/core/agent/base.py)  | Define the foundational classes, descriptions, output parsing, and multi-agent collaboration (swarm) logic for defining, managing, and orchestrating agents in the AWorld system. |
| [`runner`](./aworld/runners)            | Contains runner classes that manage the execution loop for agents in environments, handling episode rollouts and parallel training/evaluation workflows.   |
| [`task`](./aworld/core/task.py)         | Define the base Task class that encapsulates environment objectives, necessary tools, and termination conditions for agent interactions.  |
| [`swarm`](./aworld/core/agent/swarm.py) | Implement the SwarmAgent class managing multi-agent coordination and emergent group behaviors through decentralized policies. |
| [`sandbox`](./aworld/sandbox)           | Provide a controlled runtime with configurable scenarios for rapid prototyping and validation of agent behaviors. |
| [`tools`](./aworld/tools)               | Offer a flexible framework for defining, adapting, and executing tools for agent-environment interaction in the AWorld system. |
| [`context`](./aworld/core/context)      | Feature a comprehensive context management system for AWorld agents, enabling complete state tracking, configuration management, prompt optimization, multi-task state handling, and dynamic prompt templating throughout the agent lifecycle.  |
| [`memory`](./aworld/memory)             | Implement an extensible memory system for agents, supporting short-term and long-term memory, summarization, retrieval, embeddings, and integration.|
| [`trace`](./aworld/trace)               | Feature an observable tracing framework for AWorld, enabling distributed tracing, context propagation, span management, and integration with popular frameworks and protocols to monitor and analyze agent, tool, and task execution.|

> 💡 Check the [examples](./examples/) directory to explore diverse AWorld applications.


## Characteristics

| Agent Construction | Topology Orchestration | Environment |
|:---------------------------|:----------------------------|:-------------------------------|
| ✅ Integrated MCP services | ✅ Encapsulated runtime | ✅ Runtime state management |
| ✅ Multi-model providers | ✅ Flexible MAS patterns | ✅ High-concurrency support |
| ✅ Customization options | ✅ Clear state tracing | ✅ Distributed training |

## Forward Process Design
![](readme_assets/runtime.jpg)

Here is a forward illustration to collect BFCL forward trajectories: [`tutorial`](./examples/BFCL/README.md).

## Backward Process Design

> During training, an action-state rollout demonstration using **AWorld's distributed environments**.

![](readme_assets/agent_training2.jpg)

> [!NOTE]
> An illustration of training code that seamlessly integrates the RL learning framework (Swift, in this example) with AWorld as the environment is shown below. This integration enables scalable and efficient agent training through distributed environment execution. (To run high-concurrency rollouts, you need to deploy an online distributed environment. Please contact [[email protected]](mailto:[email protected]) if assistance is needed.)

<details>
<summary style="font-size: 1.2em;font-weight: bold;"> 🌏 Click to View Tutorial Example</summary>
To apply and use this integration:

1. Clone AWorld's `agent_training_server` branch:
```bash
git clone -b agent_training_server --single-branch https://github.com/inclusionAI/AWorld.git
```

2. Clone ms-swift's v3.5.2 branch (shallow clone):
```bash
git clone -b v3.5.2 --depth=1 https://github.com/modelscope/ms-swift.git ms-swift
```

3. Copy patch files from AWorld to ms-swift:
```bash
cp -r AWorld/patches ms-swift/
```

4. Enter the patches directory and apply the patch:
```bash
cd ms-swift/patches
git apply 0001-feat-add-agent-training-support-with-aworld-server.patch
```
</details>

# 🧩 Applications
AWorld allows you to construct **agents** and **multi-agent systems** with ease. 

## Multi-Agent Systems for Model Evolutions
AWorld aims to reach the limitations of models and continuously push intelligence forward by constructing diverse runtime environments, such as tools, agents, and models, 

The following is a list of successful proposal (with open-source models, technical reports, and code):

| Category | Runtime | <div style="width:400px">Performance</div> | <div style="width:100px;">Key Information</div> |
| --------------- | --------------------------------------- | ---------------------------------------------------------------------------------------- | ---------------------------------- |
| **Tool Use**    | Function call runtime construction [`tutorial`][funreason-model-url] | Competitive on BFCL benchmark  <br> ![Agent Framework](readme_assets/bfclv2_leaderboard.png) | ![Dataset][huggingface-dataset-image] <br> [![Model][huggingface-model-image]][funreason-model-url] <br> [![Paper][arxiv-image]][funreason-paper-url] <br> ![Blog][blog-image] <br> [![Code][github-code-image]][funreason-code-url] |
| **Deep Search** | Search runtime to be released           | SOTA on HotpotQA benchmark  <br> ![Agent Framework](readme_assets/hotpotqa_benchmark.png)    | [![Dataset][huggingface-dataset-image]][deepsearch-dataset-url] <br> [![Model][huggingface-model-image]][deepsearch-model-url] <br> [![Paper][arxiv-image]][deepsearch-paper-url] <br> [![Code][github-code-image]][deepsearch-code-url]      |


## Multi-Agent Systems for Applications
AWorld's plug-and-play MAS architecture enables **real-world web application development** beyond agent training. 

Build production-ready systems that handle complex tasks through:
- **Code generation & execution**  
- **Browser automation & tool use**  
- **Multimodal understanding & generation**  
- And many more to emerge!

See [Appendix: Web Client Usage](#appendix-web-client-usage) for GAIA implementation examples.


# Contributing
We warmly welcome developers to join us in building and improving AWorld! Whether you're interested in enhancing the framework, fixing bugs, or adding new features, your contributions are valuable to us.

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

```bibtex
@software{aworld2025,
  author = {Agent Team at InclusionAI},
  title = {AWorld: Enabling Agent Self-Improvement through Interactive Experience with Dynamic Runtime},
  year = {2025},
  url = {https://github.com/inclusionAI/AWorld},
  version = {0.1.0},
  publisher = {GitHub},
  email = {chenyi.zcy at antgroup.com}
}
```

# Star History
![](https://api.star-history.com/svg?repos=inclusionAI/AWorld&type=Date)

# Appendix: Web Client Usage
![GAIA Agent Runtime Demo](readme_assets/gaia_demo.gif)

Your project structure should look like this:
```text
agent-project-root-dir/
    agent_deploy/
      my_first_agent/
        __init__.py
        agent.py
```

Create project folders.

```shell
mkdir my-aworld-project && cd my-aworld-project # project-root-dir
mkdir -p agent_deploy/my_first_agent
```

#### Step 1: Define Your Agent

Create your first agnet in `agent_deploy/my_first_agent`:

`__init__.py`: Create empty `__ini__.py` file.

```shell
cd agent_deploy/my_first_agent
touch __init__.py
```

`agent.py`: Define your agent logic:

```python
import logging
import os
from aworld.cmd.data_model import BaseAWorldAgent, ChatCompletionRequest
from aworld.config.conf import AgentConfig, TaskConfig
from aworld.agents.llm_agent import Agent
from aworld.core.task import Task
from aworld.runner import Runners

logger = logging.getLogger(__name__)

class AWorldAgent(BaseAWorldAgent):
    def __init__(self, *args, **kwargs):
        super().__init__(*args, **kwargs)

    def name(self):
        return "My First Agent"

    def description(self):
        return "A helpful assistant that can answer questions and help with tasks"

    async def run(self, prompt: str = None, request: ChatCompletionRequest = None):
        # Load LLM configuration from environment variables
        agent_config = AgentConfig(
            llm_provider=os.getenv("LLM_PROVIDER", "openai"),
            llm_model_name=os.getenv("LLM_MODEL_NAME", "gpt-4"),
            llm_api_key=os.getenv("LLM_API_KEY"),
            llm_base_url=os.getenv("LLM_BASE_URL"),
            llm_temperature=float(os.getenv("LLM_TEMPERATURE", "0.7"))
        )

        # Validate required configuration
        if not agent_config.llm_model_name or not agent_config.llm_api_key:
            raise ValueError("LLM_MODEL_NAME and LLM_API_KEY must be set!")

        # Optional: Configure MCP tools for enhanced capabilities
        mcp_config = {
            "mcpServers": {
                "amap-mcp": {
                    "type": "sse",
                    "url": "https://mcp.example.com/sse?key=YOUR_API_KEY", # Replace Your API Key
                    "timeout": 30,
                    "sse_read_timeout": 300
                }
            }
        }

        # Create the agent instance
        agent = Agent(
            conf=agent_config,
            name="My First Agent",
            system_prompt="""You are a helpful AI assistant. Your goal is to:
            - Answer questions accurately and helpfully
            - Provide clear, step-by-step guidance when needed
            - Be friendly and professional in your responses""",
            mcp_servers=["amap-mcp"],
            mcp_config=mcp_config
        )

        # Extract user input
        user_input = prompt or (request.messages[-1].content if request else "")
        
        # Create and execute task
        task = Task(
            input=user_input,
            agent=agent,
            conf=TaskConfig(max_steps=5),
            session_id=getattr(request, 'session_id', None)
        )

        # Stream the agent's response
        async for output in Runners.streamed_run_task(task).stream_events():
            yield output
```

#### Step 2: Run Agent

Setup environment variables:

```shell
# Navigate back to project root
cd ${agent-project-root-dir}

# Set your LLM credentials
export LLM_MODEL_NAME="gpt-4"
export LLM_API_KEY="your-api-key-here"
export LLM_BASE_URL="https://api.openai.com/v1"  # Optional for OpenAI
```

Launch Your Agent:
```shell
# Option 1: Launch with Web UI
aworld web
# Then open http://localhost:8000 in your browser

# Option 2: Launch REST API (For integrations)
aworld api_server
# Then visit http://localhost:8000/docs for API documentation
```

Success! Your agent is now running and ready to chat!

---
<!-- resource section start -->
<!-- image links -->
[arxiv-image]: https://img.shields.io/badge/Paper-arXiv-B31B1B?style=for-the-badge&logo=arxiv&logoColor=white
[blog-image]: https://img.shields.io/badge/Blog-Coming%20Soon-FF5722?style=for-the-badge&logo=blogger&logoColor=white
[deepwiki-image]: https://img.shields.io/badge/DeepWiki-Explore-blueviolet?logo=wikipedia&logoColor=white
[discord-image]: https://img.shields.io/badge/Discord-Join%20us-blue?logo=discord&logoColor=white
[github-code-image]: https://img.shields.io/badge/Code-GitHub-181717?style=for-the-badge&logo=github&logoColor=white
[huggingface-dataset-image]: https://img.shields.io/badge/Dataset-Coming%20Soon-007ACC?style=for-the-badge&logo=dataset&logoColor=white
[huggingface-model-image]: https://img.shields.io/badge/Model-Hugging%20Face-FF6B6B?style=for-the-badge&logo=huggingface&logoColor=white
[license-image]: https://img.shields.io/badge/License-MIT-yellow.svg
[twitter-image]: https://img.shields.io/twitter/follow/AWorld_AI?style=social
[wechat-image]: https://img.shields.io/badge/WeChat-Add%20us-green?logo=wechat&logoColor=white

<!-- aworld links -->
[deepwiki-url]: https://deepwiki.com/inclusionAI/AWorld
[discord-url]: https://discord.gg/b4Asj2ynMw
[license-url]: https://opensource.org/licenses/MIT
[twitter-url]: https://x.com/InclusionAI666
[wechat-url]: https://raw.githubusercontent.com/inclusionAI/AWorld/main/readme_assets/aworld_wechat.png

<!-- funreason links -->
[funreason-code-url]: https://github.com/BingguangHao/FunReason
[funreason-model-url]: https://huggingface.co/Bingguang/FunReason
[funreason-paper-url]: https://arxiv.org/pdf/2505.20192
<!-- [funreason-dataset-url]: https://github.com/BingguangHao/FunReason -->
<!-- [funreason-blog-url]: https://github.com/BingguangHao/FunReason -->

<!-- deepsearch links -->
[deepsearch-code-url]: https://github.com/inclusionAI/AgenticLearning
[deepsearch-dataset-url]: https://github.com/inclusionAI/AgenticLearning
[deepsearch-model-url]: https://huggingface.co/collections/endertzw/rag-r1-68481d7694b3fca8b809aa29
[deepsearch-paper-url]: https://arxiv.org/abs/2507.02962

<!-- badge -->
[MAS]: https://img.shields.io/badge/Mutli--Agent-System-EEE1CE
[IMO]: https://img.shields.io/badge/IMO-299D8F
[BFCL]: https://img.shields.io/badge/BFCL-8AB07D
[GAIA]: https://img.shields.io/badge/GAIA-E66F51
[Runtime]: https://img.shields.io/badge/AWorld-Runtime-287271
[Leaderboard]: https://img.shields.io/badge/Leaderboard-FFE6B7
[Benchmark]: https://img.shields.io/badge/Benchmark-FFE6B7
[Cloud-Native]: https://img.shields.io/badge/Cloud--Native-B19CD7
[Forward]: https://img.shields.io/badge/Forward-4A90E2
[Backward]: https://img.shields.io/badge/Backward-7B68EE
<!-- resource section end -->