AutoAgent Documentation

AutoAgent is a fully-automated, self-developing framework designed to create and deploy LLM agents using only natural language. It is optimized for high performance on the GAIA benchmark and supports building tools, agents, and workflows without manual coding.

Key capabilities include:

• Agentic-RAG: Includes a native self-managing vector database.
• Universal LLM Support: Integrates with OpenAI, Anthropic, Deepseek, vLLM, Grok, Huggingface, and more.
• Flexible Interaction Modes: Supports both function-calling and ReAct interaction patterns.
• Low-Code Creation: Uses natural language to generate ready-to-use agents and workflows.

AutoAgent provides three primary modes for interacting with LLM agents:

1. user mode (Deep Research Agents): A ready-to-use multi-agent system designed for information retrieval, complex analytical tasks, and report generation. It is compatible with various LLMs (e.g., Deepseek-R1, Grok, Gemini) and supports file uploads.
2. agent editor (Agent Creation without Workflow): Allows you to create tools and agents using only natural language. The process involves inputting requirements, automated profiling, outputting agent profiles, and creating tools.
3. workflow editor (Agent Creation with Workflow): Enables the creation of agent workflows via natural language descriptions. Note that this mode does not currently support tool creation.

Start a local development server to preview the website. Most changes will reflect live in the browser without requiring a server restart.

$ yarn start

To allow AutoAgent to maintain authenticated sessions and skip repeated logins, you can provide cookie JSON files.

1. Use the Google Chrome extension "Export cookie JSON file for Puppeteer".
2. Navigate to the target website and log in manually.
3. Use the extension to export the cookies and save the resulting .json file into the autoagent/environment/cookie_json/ folder.

Generate static content into the build directory. The resulting files can be served using any static content hosting service.

$ yarn build

You can deploy the website using two different methods:

1. Using SSH: Set the USE_SSH environment variable to true.

2. Using GitHub Pages (Non-SSH): Provide your GitHub username via the GIT_USER environment variable. This will build the website and push it to the gh-pages branch.

Use agent editor to build custom agents and tools through natural language conversation. The workflow follows these steps:

1. Input Requirements: Describe the kind of agent you want to create.
2. Profiling: The system performs automated agent profiling.
3. Profiles: Review the generated agent profiles.
4. Tool Creation: Create the specific tools required for the agent.
5. Task Definition (Optional): Input specific tasks you want the agent to complete.
6. Finalize: Create the agent and proceed to the next step.

To create tools from third-party platforms (like RapidAPI), subscribe to the tools on their platform and then run the process_tool_docs.py script to add your API keys.

python process_tool_docs.py

Use workflow editor to design agent workflows using natural language. The workflow follows these steps:

1. Input Requirements: Describe the kind of workflow you want to create.
2. Profiling: The system performs automated workflow profiling.
3. Profiles: Review the generated workflow profiles.
4. Task Definition (Optional): Input specific tasks you want the workflow to complete.
5. Finalize: Create the workflow and proceed to the next step.

Note: This mode does not currently support tool creation.

You can install AutoAgent by cloning the repository and installing it in editable mode using pip. Alternatively, you can use Docker, which will automatically pull the appropriate pre-built image for your architecture.

git clone https://github.com/HKUDS/AutoAgent.git
cd AutoAgent
pip install -e .