AI Agentsagentagent-based-frameworkagentic-agiagentic-frameworkagentic-workflow
Installation
# Add to your Claude Code skills
git clone https://github.com/PathOnAIOrg/LiteWebAgent
README.md
LiteWebAgent
Disclaimer: Please note that LiteWebAgent is not affiliated with any for-profit company. This is a collaborative project from PathOnAI.org, an open-source AI research community, where Danqing Zhang (danqing.zhang.personal@gmail.com) is the main contributor and lead author of the NAACL paper. If anyone claims LiteWebAgent is affiliated with any for-profit company, please contact Danqing Zhang (danqing.zhang.personal@gmail.com) for verification.
Follow-Up Work: VisualTreeSearch
VisualTreeSearch is a production-ready system for visualizing and understanding web agent test-time scaling. It builds upon LiteWebAgent to provide an intuitive framework for researchers and users to understand tree search execution in web agents.
[2025-03-06] The 'LiteWebAgent: The Open-Source Suite for VLM-Based Web-Agent Applications' paper is now available on arXiv: https://arxiv.org/abs/2503.02950
The agent harness performance optimization system. Skills, instincts, memory, security, and research-first development for Claude Code, Codex, Opencode, Cursor and beyond.
151,568
ai-agent
ai-agents
autonomous-agent
autonomous-agents
fastapi
gpt
llm
llm-agent
llm-agents
llm-framework
web-agent
web-agents
[2025-02-28] "LiteWebAgent: The Open-Source Suite for VLM-Based Web-Agent Applications" was accepted by 2025 Annual Conference of the North American Chapter of the Association for Computational Linguistics -- System Demonstration Track (NAACL 2025).
[2025-01-09] We reused LiteWebAgent infrastructure to build LiteWebAgent Tree Search, which is set for release on 02/15.
[2025-01-01] We built a standalone evaluation suite for X-WebArena evaluation and integrated it with LiteWebAgent.
[2024-12-12] Released an async version of LiteWebAgent for improved compatibility with FastAPI AI backends.
[2024-12-11] Deployed LiteWebAgent’s frontend and backend on Vercel.
[2024-11-25] We set up a Chrome extension prototype using LiteWebAgent as an AI backend server to control the Chrome browser via Chrome DevTools Protocol.
[2024-11-01] We refactored LiteWebAgent's tree search into a new repository called LLMWebAgentTreeSearch.
[2024-10-01] Completed a major refactoring of LiteWebAgent to make it flexible for importing the package, enabling the addition of web browsing capabilities to any AI agent.
[2024-09-20] We reimplemented the paper Tree Search for Language Model Agents in the LiteWebAgent framework. Now, the search agent is capable of exploring different trajectories for accomplishing web browsing tasks and returning the most promising one. This is useful for finding the optimal path to complete complex web browsing tasks in an offline manner.
[2024-08-22] The initial version of LiteWebAgent was released, providing a robust framework for using natural language to control a web agent.
1. QuickStart
From PyPI: https://pypi.org/project/litewebagent/
pip install litewebagent
Then, a required step is to setup playwright by running
playwright install chromium
Test playwright & chromium installation by running this script
python test_installation.py
Then please create a .env file, and update your API keys:
cp .env.example .env
You are ready to go! Try FunctionCallingAgent on google.com
python examples/google_test.py
2. Development mode
(1) Installation
Set up locally
First set up virtual environment, and allow your code to be able to see 'litewebagent'
Then please create a .env file, and update your API keys:
cp .env.example .env
Test playwright & chromium installation by running this script
python3.11 test_installation.py
(2) Try different agents
use prompting-based web agent to finish some task and save the workflow
## easy case
python3.11 -m prompting_main --agent_type PromptAgent --starting_url https://www.google.com --goal 'search dining table' --plan 'search dining table' --log_folder log
## more complicated case
python3.11 -m prompting_main --agent_type PromptAgent --starting_url https://www.amazon.com/ --goal 'add a bag of dog food to the cart.' --plan 'add a bag of dog food to the cart.' --log_folder log
we also provide function-calling-based web agent
## easy case
python3.11 -m function_calling_main --agent_type FunctionCallingAgent --starting_url https://www.google.com --goal 'search dining table' --plan 'search dining table' --log_folder log
python3.11 -m function_calling_main --agent_type HighLevelPlanningAgent --starting_url https://www.google.com --goal 'search dining table' --plan 'search dining table' --log_folder log
python3.11 -m function_calling_main --agent_type ContextAwarePlanningAgent --starting_url https://www.google.com --goal 'search dining table' --plan 'search dining table' --log_folder log
## more complicated case
python3.11 -m function_calling_main --agent_type FunctionCallingAgent --starting_url https://www.amazon.com/ --goal 'add a bag of dog food to the cart.' --plan 'add a bag of dog food to the cart.' --log_folder log
We use axtree by default. Alternatively, you can provide a comma-separated string listing the desired input feature types.
python3.11 -m function_calling_main --agent_type FunctionCallingAgent --starting_url https://www.airbnb.com --goal 'set destination as San Francisco, then search the results' --plan '(1) enter the "San Francisco" as destination, (2) and click search' --log_folder log
python3.11 -m function_calling_main --agent_type FunctionCallingAgent --starting_url https://www.airbnb.com --goal 'set destination as San Francisco, then search the results' --plan '(1) enter the "San Francisco" as destination, (2) and click search' --features interactive_elements --log_folder log
python3.11 -m function_calling_main --agent_type FunctionCallingAgent --starting_url https://www.airbnb.com --goal 'set destination as San Francisco, then search the results' --plan '(1) enter the "San Francisco" as destination, (2) and click search' --features axtree,interactive_elements --log_folder log
(4) auto login
First, tell Git to ignore future changes to state.json:
git update-index --skip-worktree state.json
Then run the load_state.py script and log into the websites to enable auto-login:
python3.11 load_state.py save
(5) memory
We integrated AWM (Agent Workflow Memory) into the LiteWebAgent framework. You can follow these three steps to include induced workflows as memory for the web agent, we use 'add a bag of dog food to the cart' on amazon website as an example:
Step 2: Embed and store workflows in DB for retrieval
python3.11 memory/update_vector_store.py
Step 3: Run function calling agent with memory
python3.11 -m function_calling_main --agent_type FunctionCallingAgent --starting_url https://www.amazon.com/ --goal 'add a bag of dog food to the cart.' --workflow_memory_website amazon
(6) Use LiteWebAgent AI backend
Start the Python backend server:
python3.11 -m api.server --port 5001
3. Paper reimplementation
| Paper | Agent |
|--------------------------------------------------------------------------|--------------------------------------------------------------------------------------------------------------------------------------------------------|
| SoM (Set-of-Mark) Agent | PromptAgent |
| Mind2Web
